Anthropic announced on July 28, 2026 that an internal artificial intelligence model, Claude Mythos Preview, found real mathematical weaknesses in cryptographic algorithms that help secure the internet, including a candidate for the next generation of post-quantum encryption that human experts had studied for more than two years without success.
The findings, published alongside research papers and detailed in an Anthropic blog post titled "Discovering cryptographic weaknesses with Claude," mark one of the most striking demonstrations to date of large language models contributing to advanced mathematics. According to The New York Times, which first reported the results, the work shows how AI systems could begin to challenge core assumptions behind digital security. For the latest AI developments and deeper coverage of frontier research, readers can follow AI Buzz Wire.
Two Breakthroughs: AES and HAWK
Mythos produced two distinct results. The first was an improved attack on HAWK, a signature scheme currently in the third round of a standardization competition run by the U.S. National Institute of Standards and Technology (NIST). HAWK belongs to a class of post-quantum algorithms designed to stay secure even against future quantum computers. Human cryptographers had reviewed HAWK for more than two years, yet Mythos Preview found an improved attack in roughly 60 hours, at an API cost of about $100,000.
The second result was a new attack on a reduced version of AES-128, the world's most widely used symmetric encryption standard. The model developed a novel fingerprinting method that Anthropic calls "Mobius Bridge," which removes part of the guessing an attacker must perform. Anthropic said the technique improves on the best previously known attacks by a factor of 200 to 800. The AES attack applies only to a modified version that uses seven of the scheme's ten rounds, and the HAWK scheme is still a NIST candidate, so neither finding affects systems currently in use.
A Multi-Agent System Working Largely Alone
What makes the results notable is how little human intervention was required. According to Anthropic and reporting by The Decoder, Mythos operated semi-autonomously inside a multi-agent system, generating and testing its own hypotheses. For the HAWK attack, one agent initially tried to dismiss the approach as infeasible, but a second agent found a way to fully exploit a previously undetected symmetry in the mathematical lattice that underpins HAWK's security.
The human researcher involved had a background in theoretical computer science but was not an expert in lattice-based cryptography, Anthropic said. His role was mostly limited to project management: providing simple prompts, keeping the effort on track, and later helping verify the results. On the AES task, the model initially refused, writing that "AES-128 r5/r6 is just genuinely hard" and arguing further improvements were impossible. It only began pursuing more creative approaches after the researcher encouraged it to look for "genuinely novel ideas."
Over roughly three days, the model generated several hundred million tokens, received only three more substantive prompts, and consumed about one billion tokens in total. The run cost approximately $100,000 in API fees. Human researchers who were not cryptography experts then spent several hundred hours independently checking the conclusions.
Disclosure and a New Benchmark
Anthropic said it shared the findings in advance with the U.S. government and industry partners, and coordinated disclosure of the HAWK weakness directly with the scheme's authors. The company emphasized that because neither result breaks encryption in active deployment, the disclosures were precautionary rather than emergency responses. Mythos Preview remains unavailable to the public.
To help others evaluate the cryptanalytic abilities of language models, Anthropic collaborated with researchers from ETH Zurich, Tel Aviv University, and the University of Haifa to build a benchmark called CryptanalysisBench. The benchmark is designed to let outside teams systematically test how well AI models can find and analyze weaknesses in cryptographic constructions.
Why It Matters
The announcement arrives amid a fierce debate over how quickly advanced AI should be developed and how much autonomous authority models should be given, particularly on tasks with national security implications. Cryptography underpins nearly everything people do online, from banking and messaging to government communications. While Mythos did not break real-world encryption, it narrowed the gap between what machines and expert humans can accomplish in a field that has traditionally moved at the pace of slow, careful human analysis.
The results also raise practical questions for standardization bodies like NIST, which are currently selecting the post-quantum algorithms that will protect global communications for decades. If AI models can surface weaknesses in months that escape expert review for years, the review process itself may need to incorporate automated cryptanalysis as a routine step.
Industry observers noted that Anthropic's decision to keep Mythos Preview private, while publishing the methodology and benchmark, reflects an increasingly common balance in AI safety research: demonstrating capability without freely distributing tools that could be misused. As models grow more capable at autonomous reasoning, similar discoveries in other sensitive domains are likely to follow.
Stay Ahead of AI
Get the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
