Claude Mythos AI Finds Critical Flaw in Post-Quantum Security Candidate Experts Missed for Years

Researchers using Anthropic’s Claude Mythos AI uncovered a previously unknown mathematical weakness in a post-quantum cryptography candidate after roughly 60 hours of research, exposing a flaw that had escaped years of scrutiny by specialists and adding a new example of how frontier AI is reshaping scientific discovery.

The finding does not threaten today’s encrypted internet. The affected algorithm, HAWK, is a post-quantum digital signature candidate under evaluation by the National Institute of Standards and Technology (NIST), the U.S. agency leading efforts to prepare cryptography for quantum computing. Anthropic said the research has no practical impact on deployed systems or the widely used Advanced Encryption Standard (AES), but demonstrates that AI systems can accelerate cryptanalysis while human experts remain essential for validating discoveries.

AI Compresses Years of Cryptographic Review

Anthropic, the artificial intelligence company behind Claude, said its restricted Claude Mythos Preview model identified a previously unknown mathematical symmetry in HAWK that enabled a significantly faster key-recovery attack than researchers had previously documented.

According to the company, Claude operated inside an agentic research workflow using mathematical software, academic literature and computational experiments. Researchers provided high-level direction and periodically adjusted the investigation, but the model independently explored and refined potential attack paths.

Anthropic estimated the discovery required about 60 hours of model time and roughly $100,000 in API usage before researchers spent more than a week independently verifying the result and preparing a responsible disclosure.

The company coordinated its findings with the HAWK developers and disclosed the research through the NIST post-quantum cryptography process before publishing technical papers and demonstration code.

HAWK Security Estimates Were Significantly Reduced

HAWK is one of several digital signature schemes competing in NIST’s additional post-quantum cryptography standardization program.

The newly identified attack targets HAWK-256, the smallest parameter set under evaluation. Anthropic said the mathematical improvement reduced the estimated computational effort needed for full key recovery from about 2⁶⁴ operations to roughly 2³⁸ under the research model.

Researchers said the attack does not extend to larger HAWK parameter sets in any practical way, although maintaining the intended security margin could require substantially larger keys, reducing some of the efficiency advantages that helped distinguish HAWK from competing proposals.

Anthropic emphasized that the findings do not affect other leading lattice-based schemes, including ML-DSA, formerly known as Dilithium, or Falcon.

Reduced AES Also Receives Improved Attack

Claude Mythos also helped improve cryptanalysis against a seven-round version of AES-128, a deliberately weakened research variant of the Advanced Encryption Standard used by cryptographers to study new attack techniques.

Full AES-128, which uses 10 rounds and secures modern banking systems, cloud services and government infrastructure, was not affected.

Anthropic said the model developed a new “Möbius Bridge” fingerprinting technique that accelerated an existing meet-in-the-middle attack by removing one computational step, producing an estimated speed improvement of between 200 and 800 times under laboratory conditions.

The company said the technique remains impractical against the production version of AES and would still require computational resources far beyond realistic attack scenarios.

Human Validation Remains The Critical Step

Anthropic described the project as evidence that AI is becoming increasingly capable of generating original cryptographic research, while stressing that expert review remains indispensable.

Researchers spent hundreds of hours reproducing results, testing assumptions and confirming that the mathematical arguments held under independent verification before releasing the work publicly.

“The AI accelerated discovery,” Anthropic wrote, but researchers remained responsible for validating each result before disclosure.

The company also released technical papers describing the HAWK and AES research alongside portions of the model’s reasoning process to allow independent scrutiny by the cryptographic community.

Findings Reflect Anthropic’s Defensive AI Strategy

The discoveries were produced through Project Glasswing, Anthropic’s restricted cybersecurity initiative that gives selected researchers, technology companies and government organizations controlled access to Claude Mythos for defensive security research.

Anthropic has said the model was intentionally withheld from public release because of its advanced capabilities in vulnerability discovery, exploit development and multi-step cybersecurity tasks.

Earlier Glasswing research focused on identifying implementation flaws in widely deployed software. The latest work shifts attention toward mathematical cryptanalysis, an area traditionally dependent on years of specialized academic review.

While the findings do not require software updates or changes to existing encryption systems, they suggest frontier AI systems may increasingly help researchers evaluate cryptographic designs before they become global standards, potentially shortening the time needed to identify weaknesses while leaving human verification as the final safeguard.