Overview
- Anthropic says its Claude Mythos system autonomously found a mathematical symmetry in HAWK that effectively halved the scheme’s key strength after roughly 60 hours of automated analysis.
- The company reports the effort cost about $100,000 in compute and that human researchers then spent several hundred hours validating and refining the model’s results.
- Claude also accelerated attacks on a deliberately weakened seven‑round research version of AES by an estimated 200–800 times, but these results do not break full AES or common blockchain systems.
- Anthropic says it followed responsible disclosure by notifying HAWK’s developers and government partners, and independent community verification and follow‑up analysis are expected.
- The findings highlight the need for faster, AI‑aware cryptographic testing and new coordination and governance practices for sharing AI‑found vulnerabilities before standards are finalized.