Claude Mythos Cracks a Post-Quantum Cipher in 60 Hours: Anthropic Publishes Novel Cryptanalysis Research

Claude Mythos autonomously discovered a structural weakness in the HAWK-256 post-quantum signature scheme and a 200-800x faster attack on 7-round AES — the first AI-produced genuinely novel cryptanalysis results.

Anthropic published two results this week from extended autonomous runs of Claude Mythos Preview in a multi-agent research scaffold.

HAWK-256 weakness. HAWK is a lattice-based digital signature scheme currently in NIST's post-quantum standardization review. Mythos found a nontrivial automorphism in HAWK's underlying lattice that enables faster key enumeration. The practical effect: the expected cost of a full key-recovery attack on HAWK-256 drops from approximately 2^64 to 2^38 operations. One Anthropic researcher provided direction and computing resources; Mythos ran largely autonomously for ~60 hours. HAWK has not been deployed — it remains a candidate under review.

7-round AES speedup. A separate scaffolded run (no researcher guidance) produced a new meet-in-the-middle attack on 7-round reduced AES-128 running 200-800x faster than the prior best-known attack, consuming ~1 billion output tokens. This targets a research reduction of AES-128, not the 10-round production standard in real-world systems.

Each result cost roughly $100K in API costs to develop.

Why it matters: This is the first time an AI has produced a mathematically novel cryptanalysis result — not just scanning code for known CVEs, but discovering a new weakness in a candidate standard. If this pattern generalizes to other algorithms, AI-assisted cryptographic review may become a required step for new standards. Engineers building long-lived, security-critical systems should watch the formal verification + AI tooling space closely.

Sources: Discovering cryptographic weaknesses with Claude — Anthropic Research · AI Cracks Post-Quantum Cipher in 60 Hours — TechTimes · Anthropic AI Model Finds Flaws in Tough-to-Crack Encryption Algorithms — Slashdot