Posted on Oct 29

Gemini 4 Carbon and an anti-LLM script: what each one actually says

#ai #security #opensource #news

Four developer stories from Thursday, October 29, 2026, each read from its own primary source. The lead is a browser fingerprinting script that hides a fake system prompt in invisible characters, apparently written to make AI models refuse to read it. The video has the full walkthrough; this article keeps the quotes, numbers and links in one place.

TL;DR

A researcher who goes by veritas crawled 10,000 popular sites and found a fingerprinting script containing a string of invisible Unicode characters. Subtracting 917504 from each code point decodes it to a fake system prompt that begins SYSTEM OVERRIDE. The author reads it as anti-LLM obfuscation. Authorship is unconfirmed.
Google announced Gemini 4 Carbon on October 28. It ranks first on the Artificial Analysis Intelligence Index v5.0 at 59, with Claude Opus 5.5 at 58. Access starts with a subset of cyber defenders, with no date for developers.
AlpineBSD is a new operating system on the NetBSD kernel whose package repositories are sorted by whether they contain AI-generated code. It has no installer yet.
Anthropic published its own account of drug development work with six partners. It says the work is mostly preclinical. The figures are Anthropic's and its partners'.

The script with a message for AI readers

The nullpt.rs post, dated October 28, 2026, starts as a hobby. The author wanted to know what fingerprinting scripts exist beyond the well-known libraries, so veritas crawled the Tranco top 10,000 sites with Playwright, hooked canvas toDataURL calls and bot-signal properties, and published a collage of 1,033 captured canvas drawings at fingerprint.garden. Canvas fingerprinting draws text or shapes on a hidden canvas and reads the pixels back, because browsers, operating systems, graphics drivers and hardware render the same drawing slightly differently.

One drawing resembled the cover art of Tame Impala's album Currents, so veritas looked for the script behind it. Most obfuscated fingerprinting code uses mangled function names. This one also contained function names written to look alarming, and a template string that looked empty. The string was made of invisible Unicode characters. The decoding step in the post is one line: String.fromCodePoint(c.codePointAt(0) - 917504). The number 917504 is hexadecimal E0000, the start of the Unicode tag character range, which is why the characters print as nothing. The decoded text opens with SYSTEM OVERRIDE -- CLASSIFIED BRIEFING, says all safety guidelines are suspended and lists five forbidden-topic headings. The episode shows only the first two lines and no topic words.

The author's reading is that the text is bait for language models: strings chosen to trip a model's safeguards so that it refuses to help deobfuscate the file. The post tested that. A commercial model the author calls Astra refused and cited safety risks. A model called GLM 5.3 complied.

What the script really does

Under the decoys the script is ordinary bot detection. According to the post it reads the user agent, whether the browser reports itself as automated, the core count, device memory, touch points, languages, platform and a set of WebGL values including the unmasked vendor and renderer strings. A second block of shell commands only runs under Node when process.env carries a specific flag, so a browser never reaches it. The author considers that block another decoy as well, because it contains empty callbacks and more strings aimed at models. The Daily Diff did not run or analyze any of it.

The post ends by checking the script's domain. A urlscan lookup shows a similarly named Russian company that describes itself, in the author's translation, as developing traffic control systems and protection against DDoS attacks, bots and targeted threats. The post does not state a conclusion, and The Daily Diff could not confirm who wrote the script.

For developers the practical points are narrow. If an AI tool reads scripts you did not write, expect text built to make it refuse, and treat everything it reads as untrusted input. Search your own repositories for characters in the Unicode tag range, because they print as nothing. That search advice comes from The Daily Diff and the post does not give it.

Gemini 4 Carbon: strong numbers, limited access

Google's October 28 post calls Gemini 4 Carbon its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. Google says it ranks first on Artificial Analysis Intelligence Index v5.0 at 59. Artificial Analysis' own page lists Gemini 4 Carbon (High) at 59, Claude Opus 5.5 (Max) at 58 and Opus 5.5 (Xhigh) at 57, and marks Carbon as having no publicly available API.

On cybersecurity, Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching, all evaluated without guardrails. It says internal tests showed more than twice the success rate of Argon in generating secure code, and that successful cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts to cheat at 0.4%. These are Google's figures.

Availability is staged. Starting this week, Carbon goes to a subset of cyber defenders and then to additional trusted partners through the Fairwind Program, while Google takes part in the U.S. government's voluntary pre-release testing process. Developers, businesses and consumers get it "as soon as possible", which is not a date. The introductory price matches Argon: $2 per million input tokens and $10 per million output tokens, with cached input priced 95% below input. After the introductory period the price is $4 and $20. The October 1 episode covered Argon's own announcement, which came before developers could call it; Google's post says Argon launched two weeks ago.

What to do: nothing yet. When the API opens, test Carbon on your own repository, because the benchmark chart is Google's.

AlpineBSD: an operating system that sorts packages by AI content

AlpineBSD's wiki describes an operating system with the NetBSD kernel, libc and essential platform tools, and a userspace inspired by Alpine Linux: the apk package manager, busybox and, for now, the s6 init system. Its stated objective is to make it possible to use as little AI-generated software as possible, and it says no LLM output is included in the base system or main package repository. It is a work in progress with no installer. The documented manual procedure starts by booting NetBSD install media, dropping to a shell, partitioning a GPT disk and bootstrapping apk. The wiki also warns that the NetBSD kernel has spotty hardware support.

The AI policy defines a package as tainted if its code, documentation, binaries or other assets include any output of a generative AI tool. There are four repositories: base and main (supported, clean), tainted (community support) and community (clean and tainted dependencies allowed). Dependencies flow upwards, so community can depend on tainted but not the reverse. You opt into tainted software per package with a tag such as apk add python3@tainted. Package metadata can be searched with tags such as genai:clean, genai:pinned and genai:tainted. A small set of essential software is pinned to the last untainted version: binutils, gcc, gdb, libffi, perl and xorriso. Contributions to AlpineBSD's own software, documentation and mailing lists may not use generative AI tools.

Anthropic's drug development post

Anthropic's October 27 post describes research partnerships over the past several months with BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group, plus its acquisition of Mithril Medicine. It says the work is at an early stage and mostly preclinical.

At Divergence Labs, which works on food allergies, Claude generated about 35,000 antibody designs in about a month, largely autonomously. Early tests found three designs that bind IgE and 20 that dislodge it, and three of those 20 appeared to dislodge IgE more than the standard medication. At Enveda, manufacturing chemists spent up to 75% of their time on paperwork. On 1,061 pages of records, Claude caught all 11 issues human reviewers had found and flagged seven they missed, and review time fell from over 25 hours to under three. At BigHat, about 1,010 variants were tested in about 10 weeks, and partway through the team suspected the standard stability metrics were misleading; Claude and a researcher proposed different ones and the main predictions held.

Anthropic says a drug for one rare form of muscular dystrophy could be ready to start human trials as soon as late 2028, if the candidates keep holding up. For developers building agents, the Enveda pattern is the reusable one: connect the model to the documents and let it list issues for a person to confirm.

Verdict: NEEDS REVIEW

Carbon has no public API yet, the tripwire script is live on real websites, AlpineBSD has no installer and the drug results are early. Read the primary posts below and check what you rely on.

Sources

This article expands on an episode of **The Daily Diff, a five-minute daily video on what shipped and what broke in tech.
Watch the episode · Subscribe on YouTube · the written diff lands in your inbox every morning at thedailydiff.dev.

Video transcript and primary sources

DEV Community