July 30, 2026, (Inside AI) — Anthropic’s Claude Mythos Preview has recovered signing keys from HAWK-256, a post-quantum digital signature candidate under evaluation by the US National Institute of Standards and Technology (NIST). The AI also devised a novel attack accelerating seven-round AES-128 cracking by 200 to 800 times. Neither breakthrough threatens currently deployed systems.
The HAWK-256 result is especially significant. HAWK is the sole lattice-based scheme among nine candidates NIST advanced to its third-round post-quantum process in May 2026. Claude exploited a hidden lattice symmetry that earlier researchers flagged as theoretically vulnerable but could not practically weaponize. This reduced the expected key-recovery work factor from 2^64 to 2^38. Anthropic’s implementation recovers a functionally equivalent signing key in roughly three hours and 42 minutes on a 96-core server.
For AES, Claude developed what Anthropic calls the “Mobius Bridge,” an invariant fingerprint that eliminates a 256-way guessing step from prior attacks on seven-round AES-128. The speedup is dramatic, but the attack still demands about 2^105 chosen plaintexts, far beyond real-world feasibility. The model initially refused to attempt the AES work, insisting improvement was impossible, until researchers pushed back with blunt follow-up prompts.
The process underscores a shift in research dynamics. Mythos Preview operated largely autonomously for 60 hours on HAWK and roughly three days on AES, generating about one billion output tokens. Each project cost approximately $100,000 in API usage. Human verification took far longer: two researchers spent nearly a month confirming the AES finding. Anthropic published both results alongside technical papers, reproducibility code, and a NIST forum announcement. As of July 29, NIST still lists HAWK as a third-round candidate, though no independent reproduction of the key recovery has been publicly confirmed.
Read: Canada Regulator Warns Banks of Anthropic Claude Mythos Cyber Risks
This cryptographic milestone builds on a series of AI-driven mathematical breakthroughs. In recent months, Google used Gemini to solve open Erdős problems, OpenAI resolved the unit distance conjecture with GPT, and Anthropic’s Claude Fable 5 settled the Jacobian Conjecture. The HAWK and AES findings now extend these capabilities into practical cryptanalysis. The attacks are the strongest Claude has found to date, shared after consultation with US government and industry leaders.
Anthropic’s disclosure raises urgent questions about responsible vulnerability disclosure when AI discovers cryptanalytic breaks. The company warns that as models grow more capable, the community must prepare for scenarios where AI finds flaws in systems with immediate real-world impact. The HAWK-256 attack targets only a challenge parameter; larger variants HAWK-512 and HAWK-1024 remain impractical to break. But the speed of discovery, from theoretical weakness to functional exploit, signals a new era in security research.
Independent verification and broader implications are now under scrutiny. Meanwhile, the AES attack, though theoretical, illustrates how AI can find novel invariant structures that eluded human cryptanalysts.
Read: OpenAI Faces Competing Claims Over Navier-Stokes Maths Proof
Anthropic’s statement captures the moment: “This is not the first time that language models have performed research-level mathematics… Our result here—that Claude is able to perform cryptographic research at the level of top experts—indicates that these same capabilities also have applications in the field of cryptography, and thus may soon have more practical consequences.” The company emphasized that the attacks are the strongest found to date, shared after consultation with US government and industry leaders. As AI models continue to advance, the line between theoretical and practical vulnerabilities may blur faster than the security community expects.
