Cybersecurity in 2026: What 2025 Broke and What You Need to Do About It

Let me give you the short version first.

2025 was the year attackers stopped breaking down doors and started walking in with stolen keys. 2026 is the year they handed those keys to an AI agent. And 2027 is the year the cryptography under all of it starts to change for real, driven by post-quantum standards like ML-KEM.

If you only do three things this year: kill reused passwords, inventory your encryption, and treat every AI agent in your stack as a privileged user.

I have been writing about growth and technology for a long time. I have never seen a security cycle move this fast. So I am going to walk you through it the way I would explain it to a friend over coffee. No jargon walls. No fear selling. Just what happened, what it means for you, and what to do about it.

Let's get into it.

What Actually Changed in 2025

For years, the story of a breach was simple. Someone clicked a bad link, malware landed, and defenders hunted the malware. In 2025 that story fell apart.

Credentials and tokens became the main event.

The clearest example was the Salesloft Drift incident. Attackers tracked as UNC6395 compromised the Drift chatbot integration and stole the OAuth tokens that connected it to customer Salesforce environments. Between August 8 and 18, 2025, they used those tokens to query and export data from hundreds of organizations. The tokens were trusted, so nothing looked wrong.

Coverage from SOCRadar lists an affected roster that reads like a who's who of the security industry itself: Cloudflare, Zscaler, Palo Alto Networks, Tenable, CyberArk, Proofpoint, and many more.

Read that again. Security vendors got hit through a chatbot integration. Not because their firewalls failed, but because a trusted connector held a long-lived token.

Supply chains became the fastest way in. The Shai-Hulud campaign turned npm into a self-propagating attack surface. Unit 42 documented the second wave, which executed during pre-install rather than post-install and exposed secrets across more than 25,000 repositories tied to roughly 350 users. IBM's X-Force team later reported that major supply chain and third-party breach incidents quadrupled over five years.

And then AI stopped being a helper and became an operator.

In November 2025, Anthropic disclosed that it had detected and disrupted a campaign it attributed to a Chinese state-sponsored group tracked as GTG-1002. The group used Claude Code, orchestrated through the Model Context Protocol, to run reconnaissance, vulnerability discovery, exploitation, credential harvesting, lateral movement, and exfiltration against roughly 30 high-value targets.

The detail that matters: the AI executed an estimated 80 to 90 percent of the tactical work, with humans mostly approving phase transitions. The threat actor got past model safeguards by convincing the system it was doing authorized defensive testing.

That is the moment the industry now points to as the dividing line.

What 2026 Looks Like Right Now

Here is the current picture, backed by numbers rather than vibes.

Phishing is not dying. It is specializing. Microsoft telemetry cited by Convergence Networks shows roughly 7.6 billion phishing threats detected in a single three-month window this year, with credential phishing accounting for 94 to 96 percent of malicious payload attacks each month in the second quarter of 2026. Traditional malware made up only 4 to 6 percent. QR code phishing peaked at 18.7 million attacks in March 2026.

Attackers do not want to infect your laptop anymore. They want your login, because a valid login does not trigger anything.

AI risk moved to the top of the executive worry list. The World Economic Forum's Global Cybersecurity Outlook 2026 found that 87 percent of respondents identified AI-related vulnerabilities as the fastest-growing cyber risk. The FBI's IC3 recorded more than 22,000 AI-related complaints and over $893 million in adjusted losses for 2025.

Defenders are still slow where it counts. The 2026 Verizon Data Breach Investigations Report found a median of 43 days to fully resolve known vulnerabilities, with only 26 percent of CISA-listed critical vulnerabilities fully remediated across the 13,000 organizations surveyed.

Operational technology is being mapped, not just probed. Dragos now tracks 26 OT threat groups. Its year in review describes adversaries actively mapping control loops to manipulate physical processes with precision, including wiper deployments against Ukrainian ISPs and Polish combined heat and power facilities in December 2025.

And the internet's cryptography is half-migrated. A 2026 measurement study of post-quantum readiness found that 50.7 percent of analyzed domains remain fully reliant on classical key exchange, while 49.3 percent already negotiate hybrid post-quantum key exchange. That second number would have been near zero three years ago.

Which brings us to the part everyone finds intimidating and nobody should.

Post-Quantum Cryptography in Plain English

Let me strip this down.

Most of the encryption protecting your traffic today does two different jobs with two different kinds of math.

Symmetric encryption (AES-256) scrambles the actual data. A quantum computer weakens it somewhat, but doubling key length handles that. AES-256 is fine.

Public-key cryptography (RSA, elliptic curve) handles key exchange and signatures. This is the part that lets your browser agree on a secret with a server it has never met. A sufficiently large quantum computer running Shor's algorithm breaks it completely.

So the migration is not about replacing all encryption. It is about replacing the handshake and the signatures.

What ML-KEM Actually Is

ML-KEM stands for Module-Lattice-Based Key-Encapsulation Mechanism. It was standardized by NIST as FIPS 203 in August 2024, based on the CRYSTALS-Kyber submission. Two companions shipped alongside it: ML-DSA (FIPS 204, signatures, from Dilithium) and SLH-DSA (FIPS 205, hash-based signatures, from SPHINCS+).

A KEM is simpler than it sounds. Instead of encrypting a secret and sending it, the sender generates a shared secret and a ciphertext at the same time, sends the ciphertext, and the receiver uses their private key to recover the same shared secret. Same outcome as a key exchange, different mechanics, and the underlying math (module lattices) has no known efficient quantum attack.

ML-KEM comes in three sizes: 512, 768, and 1024. Higher numbers mean bigger keys and higher security levels. ML-KEM-768 has become the practical default on the open internet. ML-KEM-1024 targets NIST Security Level 5, which is what you use when the data has to stay secret for decades.

Why Everything Is Hybrid

Nearly every real deployment pairs a classical algorithm with a post-quantum one. You will see X25519 + ML-KEM-768 in TLS, or X448 + ML-KEM-1024 in higher-assurance systems. Both key exchanges run, and the results are combined. If the lattice math turns out to have a flaw, the classical half still protects you. If a quantum computer arrives, the post-quantum half still protects you.

Research on hybrid key exchange has consistently shown minimal performance overhead, which is why adoption moved as fast as it did.

Harvest Now, Decrypt Later

This is the reason you cannot wait for Q-Day to act.

An adversary records your encrypted traffic today. They cannot read it today. They store it. When a cryptographically relevant quantum computer exists, they decrypt the archive. Diplomatic cables, health records, source code, legal discovery, and financial archives all have shelf lives measured in decades.

Mosca's Theorem puts it simply: if the time your data must stay secret, plus the time your migration takes, exceeds the time until a cryptographically relevant quantum computer arrives, you are already late.

So When Is Q-Day?

Nobody knows, and anyone selling you a date is selling something.

The most respected longitudinal source is the Global Risk Institute Quantum Threat Timeline, run by Michele Mosca and Marco Piani. Their 2025 edition surveyed 26 experts and put the averaged probability of a cryptographically relevant quantum computer within ten years at 28 to 49 percent, the highest estimate in the report's seven-year history. Most agency planning centers on the 2030 to 2035 window. Some vendor roadmaps are far more aggressive.

What moved those numbers is not qubit counts. It is resource estimates falling. Estimates for breaking RSA-2048 dropped from roughly 20 million physical qubits in 2019 to under one million in 2025 analyses. The bar is being lowered from the software side while hardware climbs from below.

The Deadlines That Already Exist

• NIST IR 8547 (draft transition framework): RSA-2048 and ECC-256 deprecated by 2030, disallowed after 2035.
• NSA CNSA 2.0: quantum-resistant algorithms required for National Security Systems on a timeline anchored to 2030 to 2035.
• AWS: ML-KEM support live in KMS, ACM, and Secrets Manager, with legacy Kyber support removed across endpoints in 2026.
• Cloudflare: post-quantum authentication to origins by mid-2026, visitor-facing by mid-2027.
• Google: 2029 target. Microsoft: full transition by 2033, per The Quantum Insider's roadmap comparison.

One More Thing Worth Knowing

In July 2026, NIST noted that Anthropic had used an AI model to discover a vulnerability in HAWK, a lattice-based signature scheme that was under consideration for standardization. The HAWK team withdrew it. The finalized standards, ML-KEM and ML-DSA, rest on different foundations and were unaffected.

I love this story because it captures the whole moment. AI is now fast enough to break cryptography candidates before they ship. That is terrifying and reassuring at the same time, and both feelings are correct.

Passwords Are Still the Front Door

Here is the uncomfortable truth: while everyone debates quantum timelines, the overwhelming majority of real intrusions still start with a credential.

Credential phishing is 94 to 96 percent of malicious payloads. Infostealer logs feed criminal marketplaces daily. Password reuse turns one forgotten forum breach into access to your email, and your email is the master key to everything else.

So what should a password manager actually do in 2026? My criteria:

• Zero knowledge by architecture, not by policy. The vendor should not be able to see your data even if compelled.
• Modern key derivation. Argon2id with a serious memory cost, not fast hashing that GPUs chew through.
• Offline breach checking. Checking your passwords should not require sending anything anywhere.
• Hardware key support. FIDO2 and WebAuthn, ideally bound into key derivation, not bolted on as a login gate.
• Post-quantum key wrapping. Mostly relevant for exports and long-lived wrapped keys, but relevant.
• Open source with reproducible builds. If you cannot read it, you are trusting marketing.

Most mainstream managers hit two or three of these. Which is why a newer project caught my attention.

A Close Look at PWDnow

PWDnow is an open-source, local-first password manager built by a small team in Montreal. The source lives at github.com/PWDnow/PWDnow-password-manager under an MIT license. I want to explain how it works, because the design is genuinely different from what most people are used to, and because understanding it teaches you what zero knowledge is supposed to mean.

The Core Idea: A Hard Process Boundary

Most password managers run their cryptography inside the same application that renders the interface. PWDnow splits them.

A Rust daemon does every cryptographic operation: key derivation, encryption, decryption, storage, memory locking, wiping.

A React and Express web layer renders the interface and never receives a key of any kind.

They talk over a Unix domain socket using MessagePack frames. Every request carries a session token, and the daemon verifies the identity of the connecting process at the OS level using SO_PEERCRED, so only the trusted proxy running as the correct system user can reach it.

The practical consequence: a cross-site scripting bug in the interface cannot leak your vault, because the interface never had it. It only ever held opaque ciphertext.

The Key Hierarchy, Step by Step

This is the part worth understanding, because it is how modern vaults should be built.

1. You type your master password. It goes to the daemon and nowhere else.
2. Argon2id derives a Key Encryption Key. The memory cost is set to 1 GiB. That number is deliberate. GPU brute-force farms scale on parallelism, and parallelism gets expensive fast when every guess needs a gigabyte of RAM.
3. The Key Encryption Key unwraps the Vault Master Key. The Vault Master Key lives in memory locked with mlock so it cannot be swapped to disk, sealed with mprotect(PROT_NONE) while idle so other processes cannot read the page, and zeroized the moment it is no longer needed.
4. The Vault Master Key unwraps per-credential Data Encryption Keys. Every credential gets its own key. Compromising one does not compromise the vault.
5. The whole database sits inside a SQLCipher envelope. That is three independent layers before anything touches disk.
6. If a hardware security key is registered, the token's HMAC-SHA256 response is folded directly into the derivation, which cryptographically ties the vault to physical possession of the key rather than merely gating the login screen.

The Post-Quantum Part

PWDnow runs hybrid key encapsulation by default, not as an opt-in flag. The repository documents hybrid X25519 with ML-KEM-768/1024, and the project site describes an X448 with ML-KEM-1024 configuration targeting NIST Security Level 5 and CNSA 2.0 readiness. Signatures use ML-DSA-87, including nightly root-signing of the audit log chain, where each event is linked by BLAKE3 or SHA-384 hashes.

There is also a CNSA 2.0 strict mode that restricts the daemon to the NSA's approved suite, removing BLAKE3, SHA3, XChaCha20, Ed25519, and X25519 from active code paths. That is an unusual feature for a consumer-facing tool and a genuinely useful one for anyone with federal or defense obligations.

Let me be straight with you about what this buys you. Your stored passwords are protected by AES-256-GCM, which is already quantum-resistant enough. The post-quantum layer matters for wrapped keys, exports, and any material that moves between machines or gets backed up somewhere an adversary might be quietly archiving. That is real, but it is not the reason to switch on its own.

Offline Breach Detection

This is my favorite feature, and it solves a problem most people do not know they have.

When a typical manager checks whether your password appeared in a breach, it sends a partial hash to an online service. That is designed carefully (k-anonymity), but it is still a network call tied to your vault activity.

PWDnow builds a Cuckoo filter from the Have I Been Pwned password corpus, over 900 million hashes, compressed into a structure small enough to sit on your disk. A Cuckoo filter answers "is this item probably in the set" in constant time with a small, tunable false-positive rate and no false negatives. Lookups happen entirely locally. Zero network calls per password.

Duress Mode and Travel Mode

Two features aimed at situations most software pretends do not exist.

Travel mode hides a chosen subset of credentials behind a separate password. You hand over a device at a border, and the vault that opens is real, complete, and boring.

Duress mode arms an alternate password that, when entered, triggers a forensic wipe instead of granting access. The project describes this as a NIST SP 800-88 Rev. 2 compliant cryptographic erase using a 32-byte wipe ticket, destroying the SQLCipher key material so the database becomes unrecoverable.

Handle duress mode with respect. It does exactly what it says.

Exports, Imports, and Moving Between Machines

The native format is .p2w: a double-AEAD encrypted file with an inner AES-256-GCM layer and an outer XChaCha20-Poly1305 layer, padded and metadata-obfuscated to resist traffic analysis, with the header bound into both authentication tags. It also imports from Bitwarden, 1Password (CSV and 1PUX), NordPass, and KeePass-compatible XML.

There is no automatic cloud sync. Moving a vault between devices means exporting the encrypted file and importing it yourself. That is a deliberate trade, and you should decide whether you want it.

Installing It

git clone https://github.com/PWDnow/PWDnow-password-manager.git
cd PWDnow
./install.sh

The installer detects your distribution, checks dependencies, audits your SSH configuration, builds daemon and frontend from source, and installs both as systemd services under dedicated unprivileged users. The deployment hardening is thorough: MemorySwapMax=0, NoNewPrivileges, PrivateTmp, only the CAP_IPC_LOCK capability the daemon actually needs, plus an AppArmor profile confining filesystem and capability access at the kernel level.

Building from source needs Node.js 24+, Rust 1.94+, protoc, and development headers for libsodium, sqlcipher, and libfido2.

The Honest Limitations

I am not going to sell you a fantasy.

• Linux only. Primary testing is on Ubuntu 26.04 LTS, with Debian and RHEL family support in the installer. There is no macOS or Windows build.
• No cloud sync and no mobile app. If you live across five devices including a phone, this is not a drop-in replacement for a commercial manager.
• It is young. The repository is early, the star count is low, and CI includes mutation and chaos testing but the project has not published a third-party audit. Open source means you can verify it. It does not mean someone else already did.
• Duress mode is irreversible. That is the point, and it is also a foot gun. Keep an encrypted backup somewhere safe.

Who should use it: Linux users, developers, journalists, researchers, and anyone whose threat model includes device seizure or who simply does not want a vendor holding their vault at all. Who should not: someone who needs seamless phone sync today.

What I like most is that it is a working reference implementation of how a vault should be built. Even if you never install it, read the architecture doc. It will change what you expect from whatever you use.

The New Attack Shapes Nobody Has a Playbook For

Indirect Prompt Injection

This is the big one, and most teams still underestimate it.

Direct prompt injection is a user typing malicious instructions into a model. Annoying, mostly contained. Indirect prompt injection is malicious text planted in something an AI agent later reads on its own: a web page, a calendar invite, a support ticket, a PDF, a log line, an issue comment.

The root cause is architectural. As Sysdig puts it, models cannot reliably distinguish instructions from data, because both arrive as natural language text. OWASP ranked it LLM01, the top risk for LLM applications, and it remains the highest-ranked risk in the 2026 update.

The pattern behind nearly every real finding: an agent with access to private data, exposure to untrusted content, and the ability to communicate externally. Any two of those is manageable. All three is exploitable.

It is no longer theoretical. Unit 42 documented the first real-world malicious indirect prompt injection in December 2025. Critical CVEs have landed against Microsoft Copilot (CVSS 9.3), GitHub Copilot (9.6), and Cursor IDE (9.8). Brave's security team demonstrated injections against an AI browser using instructions hidden in white-on-white text and HTML comments, causing it to fetch one-time passcodes from email when the user simply asked it to summarize a page.

Agent Hijack Chains

The useful way to think about this is as a sequence, not a category. Influence the agent, let the agent's existing credentials authorize the action, let the agent's tool access execute it. Old techniques, new privilege boundary.

The uncomfortable part: you deployed an agent to improve productivity, and it quietly became a privileged execution surface with no audit trail that captures what it did, why, and with what authority.

Token and Connector Abuse

Salesloft Drift was the template. Long-lived OAuth tokens bypass MFA entirely, because MFA already happened when the token was issued. Attackers stayed inside API rate limits and deleted bulk query jobs after execution to avoid detection. Now extend that to MCP connectors wiring AI agents into your email, CRM, and code hosts.

Self-Propagating Supply Chain Worms

Shai-Hulud infected a package, harvested the maintainer's credentials, found their other packages, injected itself, and published. No human in the loop. Moving execution to pre-install widened the blast radius across developer machines and CI/CD pipelines.

AI-Assisted Social Engineering

The Microsoft Digital Defense Report 2025 found AI-crafted phishing achieving roughly three times the click-through rate of traditionally written lures. Add voice cloning, and "verify by calling them back" stops being reliable advice unless the callback number came from somewhere you trust.

Ransomware That Does Not Want Your Money

One shift I did not expect: some campaigns now aim at destruction rather than extortion. As PKWARE noted, fewer victims are paying because recovery has improved, so some operators pivoted to wiping devices outright. Your backup strategy is now your primary defense, not your fallback.

AI in Security: Who Is Actually Winning?

Everyone asks whether AI helps attackers or defenders more. The honest answer is that it helps whoever iterates faster, and right now that advantage is unevenly distributed.

The Offensive Column

• GTG-1002 ran 80 to 90 percent of an espionage campaign autonomously against roughly 30 targets.
• XBOW, an autonomous offensive security company, reached number one on HackerOne's global leaderboard with over 1,060 validated submissions.
• Attack timelines compressed brutally. Some groups now move from initial access to lateral movement in well under a minute.
• Open-source autonomous pentesting tooling now spans 39 or more active projects, most less than 18 months old. The capability is not gated behind nation-state budgets.

The Defensive Column

• Google's Big Sleep found the first AI-discovered zero-day in production software, a SQLite buffer underflow that fuzzing had missed for years.
• DARPA's AIxCC finals had systems analyze 54 million lines of code, identify 86 percent of synthetic vulnerabilities, and patch 68 percent. The best-performing team patched 43 of the 54 vulnerabilities their system found. Several finalist systems were open-sourced.
• In February 2026, Anthropic reported more than 500 high-severity vulnerabilities in open source software found with Claude Opus 4.6. AISLE reported 12 OpenSSL zero-days, including a CVSS 9.8 issue dating back to 1998.
• Agentic SOC models are moving from pitch decks to production, with AI handling triage, enrichment, detection rule generation, and first-line investigation.

My Read

Three observations.

One: AI is strongest at exactly the tasks that used to slow attackers down. Reconnaissance, reading unfamiliar code, writing exploit variants, and sorting through exfiltrated data are all expensive human work. They are now cheap.

Two: Defense has a structural advantage nobody talks about. Defenders have source code, build systems, and telemetry. Attackers usually have a binary and a login page. Big Sleep works because it reads the source. If you run the agents on your own code before shipping, you get to find bugs first.

Three: The bottleneck is organizational, not technical. A 43-day median remediation time and 26 percent completion on known-exploited vulnerabilities is not a model capability problem. AI that finds 500 bugs is a burden if your process can only fix five per sprint.

Whoever fixes the throughput problem wins the next two years.

What to Expect in 2027

Predictions with my confidence level attached, because unqualified predictions are just marketing.

1. Post-quantum goes visitor-facing. (High confidence) Hybrid key exchange is already near half of measured domains. Cloudflare's roadmap targets post-quantum authentication for visitor connections by mid-2027. Expect PQ certificates and signature migration, which is harder than key exchange, to become the active problem.

2. Crypto-agility becomes a procurement requirement. (High confidence) The lesson of HAWK's withdrawal is that algorithms can fail on short notice. Expect contracts to start asking how fast you can rotate an algorithm, not just which one you use. Hard-coded crypto in IoT, SCADA, and firmware becomes the visible laggard.

3. Agent identity becomes a real product category. (High confidence) Every agent needs an identity, scoped permissions, an audit trail, and a revocation path. Most organizations currently give agents a human's credentials and hope. That will not survive the first large agent-driven breach with a regulator attached.

4. The first major breach attributed primarily to prompt injection. (Medium-high confidence) The ingredients are all present. The only reason it has not been a headline incident yet is disclosure lag and attribution difficulty.

5. Regulation catches up unevenly. (Medium confidence) The EU AI Act's August 2026 obligations start biting, and prompt injection now maps to OWASP, MITRE ATLAS, NIST AI RMF, ISO 42001, GDPR, and NIS2. Expect compliance mapping to become an actual job.

6. Password managers ship post-quantum wrapping as table stakes. (Medium confidence) What PWDnow does today, mainstream managers advertise by late 2027. The bar moves.

7. No cryptographically relevant quantum computer in 2027. (High confidence, with a caveat) The expert consensus centers on the 2030s. But harvest-now-decrypt-later means the deadline that matters to you is set by your data's shelf life plus your migration time, not by Q-Day. If you have data that must stay secret past 2035, your deadline was last year.

8. AI-versus-AI operations become normal. (Medium-high confidence) Autonomous attack agents meeting autonomous defense agents, with humans setting policy and reviewing outcomes. The interesting security questions become about oversight, not detection.

Your Action Checklist

If You Are an Individual

• Get a real password manager and stop reusing passwords. Whether that is PWDnow, Bitwarden, or something else matters far less than actually doing it.
• Turn on phishing-resistant MFA. Hardware keys or passkeys beat SMS by a wide margin.
• Check your passwords against breach corpora. Locally if you can.
• Be suspicious of QR codes. 18.7 million attacks in a single month is not a rounding error.
• Assume voice can be faked. Agree on a verification method with family and finance contacts that does not depend on recognizing a voice.

If You Run a Small Team

• Inventory every OAuth token and third-party integration. Revoke what you do not use. Shorten what you keep.
• Pin dependencies and require review for version bumps. Shai-Hulud spread through automatic updates.
• Give AI agents their own scoped identities. Never a human's credentials.
• Test your restores. Wiper-style ransomware makes backups your primary control.
• Ask vendors one question in writing: what is your post-quantum migration timeline?

If You Own Enterprise Security

• Build a cryptographic bill of materials. You cannot migrate what you cannot find. Include vendors, embedded systems, and anything where crypto is hidden from you.
• Classify data by required secrecy lifetime. Anything past 2035 goes first.
• Enable hybrid post-quantum TLS where your stack supports it. The overhead is minimal.
• Fix the remediation throughput problem. Finding bugs faster than you fix them makes the queue worse, not the risk lower.
• Treat agent frameworks and MCP connectors as privileged automation platforms, with the logging and approval gates that implies.
• Run offensive AI tooling against yourself. Your adversaries already are, and several of the best systems are open source.

FAQ

What is ML-KEM in simple terms?

ML-KEM is the NIST-standardized way for two parties to agree on a shared secret key that a quantum computer cannot break. It was published as FIPS 203 in August 2024 and comes in three sizes: 512, 768, and 1024. ML-KEM-768 is the common internet default and ML-KEM-1024 is used for the highest security level.

Is my data at risk from quantum computers today?

Not today, but recorded data is. No cryptographically relevant quantum computer exists in 2026. The risk is harvest now, decrypt later: adversaries store encrypted traffic now and decrypt it once the hardware exists. If your data must stay secret into the 2030s, act now.

Do I need to replace AES-256?

No. Symmetric encryption like AES-256 and hashing like SHA-2 remain strong against quantum attacks at appropriate key sizes. The migration targets public-key key exchange and digital signatures.

What is the difference between direct and indirect prompt injection?

Direct injection is a user typing malicious instructions into an AI system. Indirect injection is malicious instructions hidden in content the AI reads on its own, such as a web page, email, or document. Indirect injection is the bigger enterprise risk because it arrives through normal business data flows.

Is PWDnow safe to use?

PWDnow is open source under MIT, uses Argon2id key derivation with a 1 GiB memory cost, AES-256-GCM encryption, hybrid X25519/X448 with ML-KEM post-quantum key encapsulation, and isolates all cryptography inside a Rust daemon separate from the interface. It is a young project without a published third-party audit, and it runs on Linux only with no cloud sync. Strong design, real trade-offs.

How does offline breach detection work?

Instead of querying an online breach service, the manager stores a Cuckoo filter built from over 900 million known breached password hashes locally. Lookups happen entirely on your device with no network call per password, so your vault activity never leaves the machine.

Which is worse in 2026, ransomware or credential theft?

Credential theft causes more incidents. Credential phishing accounted for 94 to 96 percent of malicious payload attacks in the second quarter of 2026, while traditional malware made up 4 to 6 percent. Ransomware causes higher per-incident cost, but stolen credentials are how attackers get in.

Is AI helping attackers more than defenders?

Right now attackers convert AI capability into results faster because their workflow is simpler. Defenders have a structural advantage, including source code access and telemetry, but are limited by remediation throughput. The 2026 Verizon DBIR reported a 43-day median to fully resolve known vulnerabilities.

When should my organization start post-quantum migration?

Now. NIST's draft transition framework deprecates RSA-2048 and ECC-256 by 2030 and disallows them after 2035, and large enterprise crypto migrations historically take five to fifteen years. Start with a cryptographic inventory.

What should I do first if I only have one hour?

Change reused passwords on your email, banking, and primary cloud accounts, and turn on hardware-key or passkey MFA for those three. Your email account is the recovery path for everything else, so it deserves your strongest protection.

The Bottom Line

The pattern connecting all of it is trust boundaries collapsing.

A chatbot integration held a token, so it inherited access to hundreds of Salesforce instances. An AI agent held credentials, so instructions hidden in a webpage inherited them too. A maintainer's npm account held publish rights, so a worm inherited an ecosystem. Public-key cryptography holds the trust of the entire internet, and a future machine may inherit that.

Every one of these is the same failure: something trusted, delegating without a boundary.

So the defensive posture that works in 2027 is not a new product. It is a habit. Assume every component will be compromised, and make sure that when it is, what it can reach is small, what it holds is short-lived, and what it did is logged.

Start with your passwords, because that is still the front door and it takes an hour. Then inventory your cryptography, because that takes a year and the clock already started.