Research · curated 28 Jul 2026
Discovering cryptographic weaknesses with Claude
First reported anthropic.com
Coverage timeline
Single-source research — first reported, latest, and curated coincide.
Why it matters
Claude Mythos Preview demonstrating that a frontier LLM can autonomously improve state-of-the-art cryptographic attacks signals that agentic AI is becoming a capable offensive research tool defenders and cryptographers must account for.
Anthropic reports that its Claude Mythos Preview model autonomously derived an improved end-to-end key-recovery attack against the HAWK-256 post-quantum signature candidate and a 200-800x speedup for an existing meet-in-the-middle attack on seven-round AES-128. Anthropic states neither result affects production systems, since HAWK is only a NIST candidate and the AES result targets a round-reduced variant, and it published implementations and write-ups of both findings.