Anthropic says its Claude Mythos Preview model has produced cryptography results that would have been difficult to ignore even if they had come from a conventional research team. The model found an improved attack on HAWK, a post-quantum signature scheme, and a new attack on a reduced version of AES, the Advanced Encryption Standard.
The company says the findings do not affect systems currently in use. That distinction matters. The HAWK result concerns a candidate in an ongoing standardization process, and the AES result applies only to a modified version using 7 of the full scheme's 10 rounds. Even so, the work shows why AI-assisted cryptanalysis is becoming a serious subject for digital security.
What Anthropic says Mythos found
According to Anthropic, Claude Mythos Preview found mathematical weaknesses in algorithms connected to modern security. One result involved HAWK, which is one of the remaining schemes in the third round of the U.S. National Institute of Standards and Technology (NIST) competition for additional post-quantum signatures.
Post-quantum signature schemes are meant to remain secure even against future quantum computers. HAWK had already been reviewed by human experts for over two years, according to the source article, before Mythos Preview found an improved attack in 60 hours.
The model's attack used a previously undetected symmetry in the mathematical lattice on which HAWK's security depends. Anthropic says the model worked semi-autonomously in a multi-agent system. In that setup, one agent first treated the idea as infeasible, while another found a way to exploit it fully.
The second result focused on a reduced version of AES-128. AES is described in the source as the world's most widely used symmetric encryption standard for digital data, and encryption protects nearly everything people do online. The Mythos finding, however, does not apply to full AES as used today. It applies to a version with fewer rounds.
Why the AES result is limited but still notable
The AES work is important because of the research process as much as the target. Anthropic says a researcher created a scaffold that let Claude generate hypotheses and test them through experiments. From there, Mythos developed a fingerprinting method Anthropic calls "Möbius Bridge."
That method removes one of the guesses an attacker would otherwise need to make. Anthropic says it improves on the best previously known attacks by a factor of 200 to 800.
The model did not begin by assuming success. It initially refused to work on the problem because it judged further improvements to be impossible, writing: "If you want a different outcome, the target has to change … AES-128 r5/r6 is just genuinely hard." The researcher then pushed the model to search for "genuinely novel ideas."
After that, the model ran for three days, producing several hundred million tokens. The source says it received only three more substantive prompts, largely to keep the work focused. One prompt read: "gain we are not looking for low hanging fruit, we want proper research to find genuinly [sic] hard findings."
The cost and human role behind the results
Both projects were expensive by ordinary API standards. The HAWK attack cost roughly $100,000 in API use. The AES run also cost about $100,000 in API fees and used roughly 1 billion tokens.
Anthropic describes the human role as limited but still necessary. In the HAWK project, the human researcher had a background in theoretical computer science but was not an expert in lattice-based cryptography, according to the source. His work was mostly project management.
For the AES result, human prompting was also small compared with the scale of the model's work. But humans still had to check the output. Anthropic says researchers who were not cryptography experts spent several hundred hours verifying the results.
That balance is central to the story. Mythos did not simply answer a single question. It worked through a difficult research space with limited guidance. At the same time, the reported discoveries still depended on scaffolding, direction, and verification by people.
No immediate break, but a clear signal
Anthropic says neither finding changes the security of systems currently in use. HAWK remains a candidate in the NIST process, not a deployed standard. The AES attack is on a reduced version of AES-128, not the full 10-round scheme.
Still, the results are a warning sign for assumptions about the pace of cryptography research. If a restricted AI model can find improved or new attacks under these conditions, then future systems may become useful tools for probing algorithms before they are standardized or deployed.
Anthropic shared the findings ahead of time with the U.S. government and industry partners. It also coordinated disclosure of the HAWK weakness with the scheme's authors. Claude Mythos Preview remains unavailable to the public.
The company has also worked with researchers from ETH Zurich, Tel Aviv University, and the University of Haifa on a benchmark called CryptanalysisBench. The goal is to let others systematically evaluate the cryptanalytic abilities of language models.
What this means for AI and security
The most cautious reading is that AI is becoming a serious research assistant for cryptanalysis, not that it has broken the internet. The source article is explicit that today's deployed systems are not affected by these findings.
The larger implication is about review. Cryptographic algorithms are trusted only after sustained scrutiny, and Mythos Preview appears to have added a new kind of scrutiny: semi-autonomous search, hypothesis testing, and attack development at very large token scale.
That does not remove the need for experts. It may increase it. If models can generate sophisticated attacks, people still need to understand, verify, disclose, and respond to them. Anthropic's results suggest that the next phase of cryptography research may involve both human specialists and AI systems pushing difficult ideas further than either could easily do alone.