# AI Model cuts post-quantum algorithm HAWK’s security in half in just 60 hours > Source: > Published: 2026-07-28 21:19:38+00:00 Anthropic published research showing that AI model Claude Mythos Preview discovered an improved attack against HAWK, a post-quantum digital signature scheme that had survived two years of National Institute of Standards and Technology (NIST) review. The model worked semi-autonomously for roughly 60 hours at an Application Programming Interface (API) cost of about $100,000, finding a nontrivial automorphism in HAWK’s lattice structure that cuts its effective key strength in half. The attack does not affect deployed standards. **How the attack was discovered** Per the research, [Mythos](https://thecoinheadlines.com/tech-and-ai/anthropics-mythos-ai-model-found-vulnerabilities-in-classified-u-s-systems-in-hours/article-23422/) worked in a multi-agent harness with occasional human guidance, conducting an extensive literature review, substantial mathematical reasoning, and computational experiments before implementing an end-to-end verification pipeline. The model found that HAWK’s lattice had a hidden symmetry (a nontrivial automorphism) that makes a faster enumeration attack possible. It’s still exponential, but you’d need to double the key size for the same security, which kills HAWK’s lead as a post-quantum cryptography (PQC) candidate. The discovery process took approximately 60 hours, with the key insight emerging from a pair of autonomous agents collaborating: one initially rejected the idea while the other found a way to fully exploit it. **The implications for post-quantum cryptography** This is cryptography research working as intended: NIST publishes candidates so flaws surface before deployment. An earlier candidate, Supersingular Isogeny Key Encapsulation (SIKE), was broken on a laptop in an hour. What is new is the speed and the price: sixty hours, a hundred thousand dollars. The lesson is not that [post-quantum cryptography](https://thecoinheadlines.com/tech-and-ai/trump-signs-quantum-eos-moves-post-quantum-crypto-deadline-to-2031/article-23226/) is unsafe. It is that no algorithm’s security is permanent, and the gap between trusted and weakened is now shorter than most migration projects. Anthropic flagged the sleeper too: many ciphers that protect real systems have received far less scrutiny than the Advanced Encryption Standard (AES), and those live in legacy appliances and vendor stacks. To this point, the standardization process is working, but the timeline for identifying critical flaws has just been compressed. **The AES breakthrough: AI finds new cryptanalytic techniques** Moving past HAWK, Mythos also managed to pull off a better attack on a reduced-round version of AES, which is pretty much the most popular symmetric cipher in the world. The attack actually hits 7 out of the 10 full rounds for AES-128, which makes the best-known meet-in-the-middle attack anywhere from 200 to 800 times more effective. The big breakthrough here was this new *“Möbius Bridge”* fingerprinting method, which drops the computational cost for the attack by a factor of 256. The discovery process was almost entirely autonomous. [Claude](https://thecoinheadlines.com/tech-and-ai/cisa-uses-anthropics-mythos-to-audit-government-code-despite-tensions/article-25114/) produced over one billion output tokens over three days, requiring just three substantive prompts from researchers, essentially encouragement to keep searching for novel ideas rather than low-hanging fruit. At first, the model didn’t want to play along, saying AES was *“just genuinely hard”* and *“there’s nothing easy to find.”* After the researchers leaned on it to *“find genuinely hard findings,”* Claude just went ahead and retooled its own agent setup until it finally stumbled onto that Möbius Bridge idea. While the attack does not break full AES and remains impractical (it would still cost hundreds of millions of dollars to implement), it demonstrates that frontier [AI models](https://thecoinheadlines.com/tech-and-ai/white-house-nears-voluntary-ai-model-standards-deal-with-openai-google-anthropic/article-24477/) can now extend a decades-long line of academic cryptanalysis. For the AI audience, the key takeaway is the autonomy: Claude generated one billion tokens, rejected its own ideas, iterated, and produced a publishable result, with human researchers spending most of their time validating the correctness, not generating insights. The bottleneck has shifted from discovery to verification.