Potheeswaran 
Anthropic disclosed on July 28, 2026, that its Claude Mythos Preview model autonomously discovered two significant cryptanalytic advances: a substantially improved attack against HAWK, a NIST-candidate post-quantum digital signature scheme, and a novel technique that accelerates attacks on round-reduced AES by up to 800 times.
Neither finding threatens deployed systems today, but both mark a turning point in how artificial intelligence is reshaping cryptographic research.
The first result targets HAWK, a lattice-based signature scheme competing in NIST’s third round of post-quantum cryptography standardization.
Anthropic’s Claude AI Cracks Post-Quantum HAWK Signature
Despite surviving two years of expert human scrutiny, Mythos identified a previously unexploited mathematical symmetry, a nontrivial automorphism within HAWK’s underlying lattice structure. This discovery halved the scheme’s effective key strength, meaning HAWK-256’s security dropped from an estimated operations to just .
The second breakthrough concerns AES, the encryption backbone of modern internet security. Working on a 7-round reduced variant of AES-128 (the full cipher uses 10 rounds), Claude engineered a technique it named the “Möbius Bridge,” a fingerprinting method that eliminated a costly guessing step in meet-in-the-middle attacks, delivering a 200-800x speed improvement over prior published research.
What makes these results notable isn’t just the math; it’s the autonomy. For the AES attack, an Anthropic researcher built an agentic scaffold and let Claude iterate largely unsupervised. Initially, the model resisted the task, insisting AES-128 was “genuinely hard” and that “there’s nothing easy to find.”
Only after researchers explicitly pushed back, sending informally worded prompts urging Claude to pursue “genuinely hard findings” rather than “low hanging fruit,” did the model pivot, ultimately generating over a billion output tokens across three days to arrive at the Möbius Bridge insight.
The HAWK attack involved greater human collaboration, with a researcher lacking specialized expertise in lattice cryptography guiding multiple Claude agents working in parallel. In one telling episode, two AI agents debated an idea; one dismissed it as infeasible while the other pursued it to a working attack, an emergent dynamic reminiscent of human research teams.
Crucially, neither attack has real-world consequences. HAWK remains an unstandardized candidate, and the AES attack applies only to a weakened, non-deployed version of the cipher. Both results represent cryptography’s adversarial review process working as intended, stress-testing algorithms before or shortly after deployment to build long-term trust.
Anthropic also released CryptanalysisBench, developed with researchers at ETH Zurich, Tel Aviv University, and the University of Haifa, providing the broader research community with a standardized way to track AI cryptanalytic capabilities over time.
For practitioners, the takeaway isn’t panic; it’s preparation. Anthropic disclosed additional preliminary findings against LEA, Serpent-128, Salsa20, and SHA-1, signaling that AI-driven cryptanalysis is scaling across cipher families, not just isolated cases.
The company explicitly warns that human validation, not discovery, is becoming the bottleneck, with researchers spending nearly a month verifying results Claude generated in a single week.
As NIST’s post-quantum migration accelerates and quantum computing timelines shrink, security teams should treat this as an early signal: algorithms considered settled may face renewed scrutiny as AI models mature.
Anthropic says it plans an academic workshop in the coming weeks to address governance questions before AI-discovered vulnerabilities hit production systems with real stakes.
About the Author
