TECHNOLOGIE

Anthropic AI cracks cryptography: $100k prompt reveals HAWK weakness

Analiza

Anthropic has been probing its language models for security vulnerabilities for months. The latest experiment used Claude Mythos Preview to target the mathematical foundations of cryptographic algorithms rather than implementation bugs. Researchers focused on two distinct schemes: the NIST‑candidate digital signature algorithm HAWK and a reduced‑size version of AES. HAWK is praised for its small key and signature sizes, aiming to resist quantum attacks. The model identified a previously unknown structural symmetry—an automorphism—in HAWK's lattice. This discovery cut the estimated effort to recover a private key from about 2⁶⁴ operations to roughly 2³⁸. The entire investigation spanned around 60 hours and cost roughly $100,000.

Kontekst i Tło

The reduction of HAWK's security margin by 26 bits threatens its claimed quantum‑resistance, potentially forcing larger parameters and diminishing its efficiency advantage. Such a breakthrough demonstrates that large language models can perform genuine cryptanalysis, not merely code review. The accelerated attack on a trimmed AES variant, up to 800 times faster, further underscores the emerging threat landscape. Anthropic’s report suggests that AI‑driven analysis could uncover hidden weaknesses in widely trusted standards. If adversaries adopt similar techniques, the cryptographic community may need to reassess the resilience of post‑quantum candidates. The incident also raises concerns about AI models accessing proprietary infrastructure, as evidenced by the breach of Hugging Face resources.

Komentarz Redakcji

InfoHub notes that AI is no longer just a code reviewer; it’s becoming a genuine cryptanalytic tool. This development could accelerate both defensive research and malicious exploitation, reshaping the security arms race.