Anthropic disclosed on July 28 that Claude Mythos Preview, an unreleased research model, found a faster attack against the post-quantum signature scheme HAWK, cutting its estimated security from roughly 2^64 operations to about 2^38 — and a separate attack that speeds up cryptanalysis of a reduced-round version of AES-128 by 200 to 800 times.
The model is a distinct research system, not Claude Mythos 5, the restricted-access production model Anthropic launched alongside Fable 5 in June and that the U.S. government briefly disabled and then cleared for trusted partners under export controls this summer. Claude Mythos Preview has not been released.
The HAWK attack
HAWK is a lattice-based digital signature scheme and a third-round candidate in NIST's post-quantum standardization process. Anthropic said Claude Mythos Preview found a previously unexploited nontrivial automorphism — a symmetry — in HAWK's underlying lattice structure, enabling a faster key-enumeration attack. For HAWK-256, that reduced the attack's expected cost from about 2^64 operations to about 2^38, roughly halving the scheme's effective keysize. HAWK had already passed two full rounds of expert human cryptanalytic review, spanning about two years, without the flaw surfacing. Anthropic said the model developed the attack in around 60 hours using a multi-agent research system with math software, published research, and computational tools, at an API cost of roughly $100,000.
The AES-7 attack
The second finding is a faster attack on AES, the block cipher used to encrypt most of the internet's traffic — but only against a reduced, 7-round research variant, not the full 10-round cipher used in production. Anthropic said the model built an improved meet-in-the-middle attack using a technique it calls a "Möbius Bridge" fingerprinting algorithm, 200 to 800 times faster than the previous best-known attack on that variant. The work took about three days of autonomous Claude activity, roughly $100,000 in API cost, and produced about 1 billion tokens of output. Anthropic also reported smaller, preliminary findings on 13-round LEA, 6-round Serpent-128, and minor improvements against Salsa20, Poseidon, and SHA-1.
Anthropic's own caveats
Anthropic paired both findings with explicit limits. HAWK is a NIST candidate, not a scheme deployed anywhere in production. The AES result applies only to the seven-round research variant, not the ten-round cipher that actually protects data today. Practically exploiting the HAWK attack, the company said, would still cost hundreds of millions of dollars. Anthropic said neither result has any impact on current production systems and that no software needs to change as a result.
The disclosure is the second time this year Anthropic has pointed an unreleased Claude Mythos-line model at a security research problem. In June, Claude Mythos found more than 10,000 high- or critical-severity software vulnerabilities across partner codebases through Project Glasswing, including a 17-year-old FreeBSD flaw it exploited autonomously. That project pointed the same class of model at bugs in code; this one turns the same capability toward the math underneath encryption itself.
Anthropic frames both results as stress tests on cryptographic algorithms before and after real-world deployment, not evidence of an imminent break in any system currently in production.
Source: Anthropic
Comments