Researchers at Anthropic have used their own AI model, Claude Mythos Preview, to uncover previously unknown weaknesses in cryptographic algorithms — not just the code that implements them, but the underlying mathematics itself. The findings, described in two new research papers, represent a notable step forward in cryptanalysis and offer an early glimpse of how powerful artificial intelligence models could transform the way security experts hunt for flaws in the systems that protect online communications.
The first result is an improved attack against HAWK, a digital signature scheme designed to withstand future quantum computers. HAWK is currently a third-round candidate in a National Institute of Standards and Technology competition to select post-quantum cryptographic standards, and it has already survived two years of expert human review. Working with Claude over roughly 60 hours, an Anthropic researcher found a way to effectively cut the scheme’s key strength in half. The second finding targets a weakened version of AES, the symmetric cipher that secures everything from web traffic to banking data. Claude discovered an approach that speeds up previous best attacks on this reduced variant by a factor of 200 to 800.
Neither breakthrough poses any danger to systems people use today. HAWK has not been deployed in production software, and the AES attack works only on a deliberately simplified version of the algorithm rather than the full cipher. But the results are significant because they demonstrate that a frontier AI model can autonomously reason about deep mathematical structures in ways that lead to genuine cryptanalytic advances. In the case of the AES work, Claude operated almost entirely without human intervention after an Anthropic researcher built a scaffolding framework for it to work within. Each discovery cost roughly $100,000 in computing resources.
Anthropic says it followed responsible disclosure procedures throughout the process, sharing its findings with the designers of HAWK in June and coordinating public release with a notification to a NIST mailing list. The company also consulted with academic partners at ETH Zurich, Tel Aviv University, and TU Berlin to verify the results. To encourage further research into AI-assisted codebreaking, Anthropic collaborated with those same institutions to release CryptanalysisBench, a benchmark designed to help other researchers evaluate how well large language models can analyze and break ciphers. The broader message from the research is that stress-testing algorithms before they reach widespread deployment remains one of the most effective ways to build trust in digital infrastructure — and AI may soon be one of the most important tools for doing exactly that.
