Notes on the Anthropic cryptographic blogpost

LessWrong is reporting that Anthropic released research showing how Claude Mythos Preview discovered improved attacks against cryptographic algorithms. The findings include two separate results: one against HAWK, a candidate for post-quantum cryptography standardization by NIST, and another against a weakened version of AES, the widely-used encryption standard. The HAWK attack would require doubling key sizes to maintain the same security level, substantially weakening its advantages as a post-quantum candidate. The AES work attacked a seven-round version of the algorithm, a common academic approach to understanding vulnerabilities, and produced attacks between two hundred and eight hundred times faster than previous research. Both discoveries used multi-agent AI workflows in specialized harnesses, allowing Mythos to work largely autonomously with guidance from human researchers. The HAWK attack took roughly sixty hours and the AES work about a week, each costing around one hundred thousand dollars in API expenses. Neither attack threatens current production systems, but the research underscores how rapidly AI is advancing in specialized technical domains.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton