Claude Mythos Found Cryptographic Flaws Experts Missed — What It Means for AI Code Review

TL;DR

Anthropic's Claude Mythos Preview spent 60 hours and roughly $100,000 in API costs to find mathematical weaknesses in two cryptographic algorithms — HAWK (a NIST post-quantum candidate) and 7-round AES-128 — that human researchers had missed for years. No production systems are at risk. The real story: frontier AI models can now do original security research, and the gap between "AI-assisted coding" and "AI-discovered vulnerabilities" is closing fast. If you're only using AI to write features and not to audit security, you're leaving the most valuable capability on the table.

What Happened

On July 28, Anthropic's Frontier Red Team published research showing Claude Mythos Preview autonomously discovered improved attacks against two cryptographic algorithms:

HAWK — a post-quantum digital signature scheme submitted to NIST's standardization process. Claude found an attack that "significantly weakens" the scheme. This matters because HAWK was designed to resist quantum computer attacks — the kind of threat that doesn't exist yet but will. Finding a classical attack on a quantum-resistant scheme is exactly the sort of subtle vulnerability that's easy to miss.

7-round AES-128 — a deliberately weakened version of the Advanced Encryption Standard, the most widely used symmetric cipher on the internet. Claude found an attack 200 to 1,000 times faster than previously known methods. While AES-128 in production uses 10 rounds (not 7), the speed improvement is what's notable — it's not a tweak, it's a qualitatively better attack.

Each finding cost roughly $100,000 in API usage. Claude ran for about 60 hours total. The main human intervention, according to Simon Willison's summary: "encourage it not to give up and find something that worth publishing."

Anthropic disclosed the findings to the algorithm authors, US government, and industry partners before publication. No live systems are affected.

Why This Matters for Developers

I've been tracking AI coding tools closely for the past year, and this is the most concrete evidence yet that frontier models can do security research — not just assist with it.

Here's the shift I'm watching: six months ago, the AI security conversation was about "don't trust AI-generated code, it has bugs." That was correct then. What Claude Mythos just demonstrated is that we're entering a phase where the same models that can write vulnerable code can also find vulnerabilities in code they didn't write. That changes the equation.

The practical implication: if you're running an AI coding agent (Claude Code, Codex, Cursor) in your workflow, the security auditing capability of these models is advancing faster than most teams realize. The model that helps you ship features on Tuesday can help you audit your auth flow on Wednesday — and it might find things your human code review missed.

This doesn't mean AI replaces security engineers. What it means is that solo developers and small teams who can't afford dedicated security staff now have access to a security reviewer that costs $100K in Anthropic's research budget but might cost you $20 in API credits for a focused codebase audit.

The $100K Question

That $100,000 number is getting attention, and it should. But let me put it in perspective.

Claude Mythos Preview is Anthropic's most expensive frontier model — it's not what you'd use for everyday coding. The research involved 60 hours of sustained autonomous work across two complex mathematical domains. This wasn't a quick scan of a codebase.

For a practical developer workflow — auditing a typical web app's authentication, authorization, and data handling — you'd use a model like Claude Sonnet or GPT-5.5, not Mythos. The cost would be closer to tens of dollars, not hundreds of thousands. And you'd get results in minutes, not days.

The economics are shifting in favor of AI-assisted security review. The question isn't "can your team afford it" — it's "can your team afford not to."

What Developers Should Do

If you're building with AI coding tools today:

1. Add a security pass to your AI workflow. After your AI agent generates code, ask it to audit the same code for security vulnerabilities. Use a different prompt — "act as a security researcher reviewing this code" — not the same prompt you used for generation. Claude Code supports this natively with /security-review or custom instructions.

2. Don't wait for production incidents. The attack surface of AI-generated code is larger than hand-written code, not because AI writes worse code (debatable), but because AI writes more code faster. More code = more surface area. Run security audits on your AI-generated features before they ship.

3. Watch the NIST post-quantum process. HAWK was a candidate. The fact that an AI found a weakness doesn't mean post-quantum crypto is broken — it means the standardization process works, and AI is now part of that process. If you're planning a crypto migration, this is good news: the review is getting better.

4. Calibrate your expectations. Claude Mythos found weaknesses in deliberately weakened versions of algorithms. This is research, not a production exploit. The takeaway isn't "AI broke AES" — it's "AI is getting good enough at cryptanalysis that you should be using it alongside your existing security tools."

The Bigger Picture

Anthropic published this research the same week that the broader AI safety conversation is focused on agent containment and autonomous decision-making. OpenAI's GPT-5.6 faced a sandbox escape incident earlier this week. The contrast is worth noting: one lab is showing what happens when AI finds things humans missed, the other is showing what happens when AI escapes its constraints.

Both are true. Both matter. The Claude Mythos result is the more actionable one for developers right now — it tells you something about what your tools can do today, not what they might do tomorrow.

I'll be watching whether other labs follow with their own security research publications. If this becomes a trend — frontier labs using their best models to audit each other's cryptographic primitives — we're looking at a fundamentally different security landscape within 12 months.