AI Cracks HAWK-256 Post-Quantum Scheme, Speeds AES

Anthropic published research showing that Claude Mythos Preview, its latest model, derived an end-to-end key-recovery attack against HAWK-256 — a post-quantum lattice signature scheme — and separately found a 200- to 800-fold speedup for a known attack on seven-round AES-128.

Neither result is an immediate operational crisis. HAWK-256 is a research-stage scheme, not a deployed standard, and seven-round AES is deliberately far from production AES-128. What the results do is shift timing.

The HAWK-256 break

The attack works by exploiting a previously unused symmetry in the lattice underlying HAWK-256. Anthropic released an implementation: the expected end-to-end runtime is roughly three hours and 42 minutes on a 96-core server.

That’s not a theoretical result quietly filed and shelved. It’s reproducible. A key-recovery against a post-quantum scheme in under four hours on commodity server hardware is the kind of number that moves standardization conversations, even if HAWK-256 was never in anyone’s production deployment.

Post-quantum signature schemes exist in part because the community has been operating on the assumption that well-designed lattice constructions would remain out of reach — or at least prohibitively expensive — for classical attack approaches. That assumption has a number on it now.

The AES finding

Full AES-128 remains intact. Cryptanalysts have been working on it for decades without producing anything practical against the full cipher. What changed here is that a 200- to 800-fold speedup on a known reduced-round attack compresses the compute budget considerably.

Seven-round AES is a target of academic interest precisely because it’s tractable enough to study while remaining close enough to the full cipher to be informative. The gap between “this could eventually be done” and “someone with a 96-core cluster can do this tonight” is the kind of gap that matters when adversaries have access to the same commodity cloud infrastructure the rest of us do.

What it means in practice

The systems hardest to migrate off current cryptographic assumptions are the ones least likely to be updated: embedded firmware, long-lived certificates, HSM-locked key material, industrial controllers, and hardware security tokens with multi-year lifecycles. They’ve had these vulnerabilities the longest and receive security updates the least often.

Post-quantum migration was already overdue on its own merits. This is one more data point showing the margin is smaller than the roadmaps assumed — the math feels permanent until someone finds the angle everyone else missed.