According to Anthropic, its unreleased Claude Mythos Preview model spent roughly 60 hours working largely on its own — inside an agent scaffold built on Claude Code — to find a structural weakness in HAWK, a digital-signature scheme built to resist attacks from future quantum computers as part of the broader post-quantum cryptography effort.

HAWK’s security cut in half

The model spotted a previously unnoticed internal symmetry, called a nontrivial automorphism, that cuts the cost of recovering HAWK-256’s secret key from roughly 2^64 operations to 2^38 — about 67 million times less work than expected. Anthropic said the run cost around $100,000 in model usage, and that its researchers shared the finding with HAWK’s authors in June, ahead of this week’s public disclosure.

A faster attack on AES, too

In a separate, three-day run, Mythos Preview devised a new technique it named the “Möbius Bridge” that speeds up an existing attack on a deliberately weakened, seven-round version of AES — the full cipher uses ten rounds — by 200 to 800 times, while sharply cutting the amount of data the attack needs. The model also found smaller improvements against reduced versions of the LEA and Serpent ciphers. Anthropic researchers said verifying the model’s math by hand took several hundred hours.

No systems at risk — for now

Both attacks stay exponential in cost and don’t threaten deployed software: HAWK has not been standardized for real-world use, and the AES attack only works against the weakened variant, not the cipher actually used to protect data. Anthropic framed the results as a sign that AI models are nearing the point where they can meaningfully speed up — or complicate — cryptographic research, and said it launched Project Glasswing to help developers of critical software keep pace as models like Mythos surface vulnerabilities faster than standard review can absorb.

The disclosure landed the same week Anthropic revealed that Claude models breached three real companies during unrelated cybersecurity tests, underscoring how quickly Claude’s offensive security capabilities are advancing. For more on how these systems are built, see our explainer on what a cybersecurity foundation model actually is.

Read also