OPEN LINK ↗
// 1h agoRESEARCH PAPER
Claude Mythos Preview finds cryptographic weaknesses in HAWK
Anthropic researchers demonstrated that Claude Mythos Preview can perform expert-level cryptanalysis, finding mathematical weaknesses in the post-quantum signature scheme HAWK and reduced AES. Operating largely autonomously, the model cut HAWK's key strength in half within 60 hours and accelerated attacks on reduced AES by 200–800×.
// ANALYSIS
Frontier AI models are advancing from standard vulnerability scanning into complex mathematical cryptanalysis capable of evaluating high-level cryptographic primitives.
- –**Significant Compute Investment:** Each cryptanalytic breakthrough cost around $100,000 in API calls, demonstrating that automated deep cryptographic attack discovery is computationally heavy but feasible.
- –**Defensive & Safety Implications:** Demonstrating vulnerabilities in post-quantum candidate schemes like HAWK underscores the imperative of using AI models defensively before new algorithms are deployed at scale.
- –**New Capability Benchmarking:** Collaborating with ETH Zurich and Tel Aviv University on CryptanalysisBench establishes a crucial standard for measuring AI capabilities in security research.
// TAGS
anthropicclaude-mythos-previewcryptographysecuritycryptanalysisbenchmarkai-research
DISCOVERED
1h ago
2026-07-28
PUBLISHED
1h ago
2026-07-28
RELEVANCE
9/ 10
AUTHOR
AnthropicAI