Anthropic says Claude Mythos Preview broke weakened AES and improved a HAWK attack
Anthropic reports that Claude Mythos Preview broke a weaker version of AES and orchestrated an improved attack on the HAWK cryptographic system in internal testing. The result points to stronger autonomous cryptanalytic reasoning than typical LLM demos, with clear implications for AI-assisted security research.
Key takeaway
The story matters less as a claim that production AES is dead and more as evidence that frontier models are becoming useful collaborators for finding cryptographic weaknesses under controlled conditions.
Context
According to the Techmeme-linked report, Anthropic says Claude Mythos Preview succeeded against a weaker AES variant and also drove an improved attack on HAWK during internal evaluation. That framing keeps the result in research territory while still marking a step beyond generic coding or summarization benchmarks.
For security teams, the near-term takeaway is workflow change: models may help propose and refine cryptanalytic ideas faster, which can speed defensive audits and also raise the bar for how quickly weak designs get probed. The channel-worthy angle is the official lab attribution plus named crypto targets, not a vague claim of superhuman hacking.