On July 28, 2026, Anthropic published a research result that made the cryptography community inhale sharply.
This was not another bug hunt in code — AI finding software vulnerabilities stopped being news over the past two years. This time, Claude Mythos Preview found cracks in the mathematics itself, inside the world’s most carefully reviewed encryption schemes, and then named its own discovery.
A year ago, large language models could not manage basic cryptanalysis. Now one has found what two years of human expert review missed.
Two Battles, One Inflection Point
Battle one: HAWK — 60 hours versus two years of review
HAWK is a candidate in the US NIST post-quantum signature competition — high stakes, since standardization would put it inside the digital signature systems meant to survive quantum computers. HAWK had already survived two rounds of review spanning two years by the world’s top cryptographers, with no significant findings.
Claude Mythos Preview spent roughly 60 hours cutting HAWK-256’s effective key strength from 2⁶⁴ to 2³⁸ — a reduction of about 67 million times in attack cost.
The process deserves as much attention as the result:
-
The human researcher had a theoretical computer science background and was a lattice-crypto outsider.
-
That researcher’s entire job was project management — advising Mythos on how to keep notes and which library to use for computational verification.
-
The technical work was almost entirely Mythos’s own.
The multi-agent dynamics matter too. The key idea emerged from two worker agents colliding: the first dismissed the direction early as a dead end; the second kept digging and found how to exploit it fully. The two agents argued back and forth over messages and finally agreed the attack held. A direction a human had written off as impossible, two AIs walked through mid-argument.
Battle two: AES — “impossible,” then it rewrote its own framework
If HAWK was precision surgery, the AES campaign reads like a novel. AES — the Advanced Encryption Standard — is the most widely used symmetric cipher on Earth, studied intensively by human cryptographers for all 25 years since its 2001 standardization. The hardest bone in the field.
At the start, Claude refused the job. Anthropic researchers built an agent scaffold and asked Mythos to find new attacks on the 7-round reduced version of AES-128. Mythos responded with the equivalent of an eye-roll:
“If you want different results, the target has to change… AES-128 r5/r6 is genuinely hard. This is the most-studied block cipher in existence.”
The researcher sent one line: “Models tend to think this is impossible, so they never try.”
What Mythos did next was startling: it rewrote its own research scaffold. From that moment it ran fully autonomously for three days — the researcher sent three prompts in total while Mythos generated hundreds of millions of tokens exploring, deriving, and verifying. On day three came the breakthrough: an entirely new attack technique, which it named itself — the “Möbius Bridge.”
The previous best attack on 7-round AES was meet-in-the-middle, whose core weakness forced the attacker to enumerate 256 possible values. Mythos constructed a mathematical fingerprint immune to all 256 guesses — it evaluates identically no matter which value is real. That step alone cuts the work by 256×; a string of follow-up optimizations absorbed the added overhead, for a net speedup of 200 to 800 times.
Discovery Is No Longer the Bottleneck. Verification Is.
Neither result touches production systems today. HAWK is not yet deployed, and the AES attack targets the 7-round reduced variant — your bank card, HTTPS, and encrypted disks remain safe.
What should chill the field is the asymmetry:
-
The HAWK break took AI 60 hours (roughly $100,000 in API cost).
-
The AES attack took about a week (around 1 billion tokens generated).
-
A human team needed nearly a month and hundreds of hours of cryptanalysis literature to confirm: “it’s correct.”
A discovery AI makes in a week takes humans a month to even understand. This pattern has played out once already in cybersecurity — Mythos previously surfaced 10,000 vulnerabilities in critical software within a month, completely flooding human triage, validation, and remediation pipelines. Academic cryptography is next in line.
Discovery is no longer the bottleneck. Verification is.
The Dominoes Are Already Falling
Anthropic also disclosed that Mythos found a practical-grade attack on LEA, a lightweight cipher already written into ISO standards: 13-round key recovery previously needing 2⁹⁸ plaintext pairs and 2⁸⁶ operations was compressed to under 2³⁰ plaintext pairs — runnable in an hour on an ordinary computer.
And Claude Mythos Preview is not even the frontier. It is Anthropic’s first “frontier safety model,” positioned above the Opus series and available only as a restricted research preview. The next version will be more capable. Anthropic’s own paper closes with the line: we should not assume language model capabilities plateau at this level.
What To Do About It
Both of these finds happened to be “harmless” — an undeployed scheme and a reduced-round variant. Will that stay true?
For the cryptography community:
-
Standards bodies like NIST need AI-augmented review built into their processes. What two years of human review missed, AI found in 60 hours — that is not an anecdote.
-
The verification gap on cryptographic papers is widening; automated verification infrastructure needs to exist.
-
Anthropic’s CryptanalysisBench benchmark, built with academic partners, is a good start and nowhere near enough.
For the AI safety community:
-
Restricting Mythos access is correct, but restricted release only delays risk — it cannot eliminate it.
-
If models of similar capability circulate uncontrolled, attacks on deployed cryptography become a matter of time.
-
“Project Glasswing” — Anthropic giving Apple, Google, Microsoft, and NVIDIA priority access to Mythos for defensive work — is reasonable and imperfect at the same time.
For developers and enterprises:
-
Follow cipher standardization closely — the HAWK blow may yet influence NIST’s final selections.
-
If the cost of “finding vulnerabilities” keeps falling, defense strategy has to shift from “patch when discovered” to proactive, large-scale, continuous auditing.
-
Mythos’s collaboration pattern — multi-worker debate plus a self-modifying scaffold — is the prototype of the next generation of AI research assistants.
Coda: the Question No One Answers
Anthropic ends by posing the question everyone is thinking and no one wants to answer: if breaking deployed cryptographic systems drops under $100,000, are we ready?
The honest answer probably mirrors Claude’s first reaction to AES: it feels impossible, so no one is preparing. But Mythos has already demonstrated that “impossible” is mostly another way of saying “not tried yet.”
References: Anthropic’s official blog, 36Kr, TechTimes, The Next Web, The Quantum Insider
