Claude Found Real Security Flaws in Encryption. Here's What Professionals Need to Know.

TL;DR. On July 28, 2026, Anthropic published research showing that Claude Mythos Preview — its restricted frontier model — autonomously discovered two genuine cryptographic weaknesses: a mathematical flaw in HAWK (a post-quantum encryption candidate) found in 60 hours, and a 200–800× improvement on attacks against a research version of AES. Neither finding affects production systems today. But the episode signals something larger: AI is now doing expert-level frontier research, not just assisting with it.

Every few years, an AI milestone arrives that's harder to categorize than "it beat a human at chess" or "it passed a medical board exam." The latest one — quietly dropped by Anthropic on Monday — is that kind.

Claude Mythos Preview, Anthropic's restricted research model, was set to work on a specific type of mathematical problem: find weaknesses in cryptographic algorithms. It succeeded. Twice. The algorithms were real. The weaknesses were novel. And a team of human cryptographers needed a month to fully verify what the model produced in three days.

This is worth understanding carefully — not because it means your passwords are compromised, but because of what it tells you about where AI capabilities are heading.

What happened

Anthropic's research team gave Claude Mythos Preview two separate cryptanalysis challenges.

Challenge 1: HAWK. HAWK is a digital signature scheme — a type of cryptography used to verify identity and authenticate data. It's been in NIST's post-quantum standardization pipeline (the effort to build encryption that can withstand future quantum computers) and survived two rounds of expert peer review over two years. Mythos Preview was given minimal direction and told to find weaknesses.

In approximately 60 hours — at a cost of around $100,000 in API compute — the model found one. It identified a previously undetected symmetry in the mathematical lattice that HAWK's security relies on, and used it to reduce the effective key strength of HAWK-256 from approximately 2⁶⁴ operations (computationally infeasible) to 2³⁸ (substantially more tractable). Anthropic coordinated responsible disclosure with NIST and HAWK's authors before publishing.

Challenge 2: 7-round AES. AES (Advanced Encryption Standard) is the world's most widely used symmetric encryption. Full AES-128 runs through 10 rounds of processing. Researchers study "reduced-round" variants — fewer rounds, same structure — as a way to probe the algorithm's limits. The 7-round variant hadn't seen a meaningful improvement in attack efficiency since 2013.

Mythos Preview initially refused the task. "If you want a different outcome, the target has to change," it said. "AES-128 r5/r6 is just genuinely hard." Over three days and three encouraging prompts, researchers persisted. The model eventually produced what Anthropic called over one billion output tokens of autonomous work — and invented a novel technique called the "Möbius Bridge," which improved attack efficiency against 7-round AES by 200–800× compared to the best prior method.

The research team needed a month to verify and validate the output.

What it doesn't mean

To be clear about the security picture:

Production AES is unaffected. The 7-round variant the model attacked is a research construct, not what's protecting your bank account, VPN, or encrypted email. Full AES-128 runs 10 rounds. The gap between a 7-round attack and a 10-round attack is enormous — the additional rounds are specifically designed to defeat this class of attack. Cryptographers confirmed there is no known path from Mythos's finding to breaking production AES.

HAWK isn't deployed. It's still a standardization candidate. It has never been used to protect real data. The finding may influence whether NIST continues with HAWK as a post-quantum standard, but it has no effect on any currently deployed system.

Neither finding runs in polynomial time. The attacks are exponential — they still require enormous computation. They're faster than before, but not fast enough to be practical against secure key sizes in use today.

The threat to your data as you read this is unchanged from yesterday.

What it does mean

The security story is relatively contained. The capability story is not.

Before this research was published, you could reasonably argue that AI was a useful assistant for cryptographic work — helping researchers find references, write code, organize literature, spot patterns in known attacks. What Anthropic published changes that framing. Claude Mythos Preview didn't assist human cryptographers. It did the research itself, autonomously, with three prompts over three days. The human team provided project direction and, afterward, a month of verification work.

The 60-hour HAWK finding is especially striking. Two rounds of NIST peer review, years of expert scrutiny by professional cryptographers, and a 60-hour AI run found something they didn't. NIST's peer review process is rigorous — this isn't a knock on the human reviewers. It's a statement about what AI can do when deployed at scale against a well-defined search problem.

The pattern this fits

Cryptanalysis has specific properties that make it tractable for AI: there's a clear objective (improve the attack), a measurable outcome (operations required), and a known solution space (attack families from the literature). AI performs particularly well when the problem has those characteristics.

But those characteristics also describe a large number of expert knowledge domains. Due diligence research. Drug interaction screening. Regulatory compliance gaps. Contract analysis for edge cases. Pattern recognition across large case law databases. The specific mechanisms differ from cryptography, but the structure — large search space, well-defined success criteria, existing human literature as a starting point — is similar.

Anthropic didn't publish this research to show that AI is smarter than cryptographers. They published it as a demonstration of what frontier AI can do when applied autonomously to expert problems. The honest read is that cryptanalysis was a proof-of-concept domain. The actual territory is much broader.

What changes for professionals using AI tools

In the near term, nothing about what AI tools you use or how you use them. Mythos Preview is not available to most users; Fable 5 and Claude Opus 5 remain the ceiling for non-Glasswing organizations. The research result doesn't affect any tool's terms of service or privacy posture.

The longer-arc implication matters more. What Claude Mythos Preview demonstrated is that AI at the frontier can do autonomous, multi-day expert research work and produce novel findings. That capability doesn't stay at the frontier indefinitely — it filters into more accessible models over time, the same way image generation, code completion, and multi-document reasoning all moved down the capability ladder over the past four years.

If AI can independently produce frontier cryptographic research today, it will produce less-frontier but still-expert-level analytical work in your field sooner than most planning assumptions account for. The practical shift for professionals isn't switching tools next week. It's adjusting expectations about which parts of expert knowledge work remain human-requiring over a 2–5 year horizon.

That's not cause for alarm — it's cause for thinking ahead. The professionals who benefit most from AI aren't the ones who treat each capability milestone as a news item. They're the ones who connect it to the actual shape of their work.

---

Sources

• Anthropic Research: Discovering Cryptographic Weaknesses with Claude (July 28, 2026) — primary source, full technical details
• The Decoder: Anthropic says its Mythos model found vulnerabilities in cryptographic algorithms that secure the internet (July 28, 2026)
• The Next Web: Claude found mathematical flaws in two cryptographic algorithms that years of expert review missed (July 28, 2026)
• AI Weekly: Anthropic Says Claude Mythos Found HAWK, 7-Round AES Flaws (July 28, 2026)