AI-assisted cryptanalysis knocks HAWK out of the post-quantum race

Lead

A post-quantum cryptography candidate that was still being considered for a U.S. standard has effectively been removed from contention after an AI-assisted discovery exposed a serious weakness. According to Ars Technica, Anthropic’s security model Mythos helped researchers improve an attack on HAWK, a digital signature scheme designed to resist future quantum computers. Shortly after the disclosure, HAWK’s developer said he was withdrawing it.

The finding is notable because HAWK was not an obscure toy system. It had already survived two rounds of review in NIST’s post-quantum cryptography process and was undergoing a third round intended to catch deeper flaws. That is exactly the kind of review phase in which this result now lands.

Key points

• HAWK was weakened, not a deployed system broken. Anthropic says a researcher without cryptography expertise used Mythos for about 60 hours and roughly $100,000 in compute to improve the best-known attack on HAWK, effectively cutting its key strength in half.
• The attack recombined known ideas. HAWK relies on the hardness of the Lattice Isomorphism Problem. The reported advance involved a new way to find automorphism symmetries, building on tools that were already known but had not been put together in this way.
• Mitigation would reduce HAWK’s appeal. Doubling the key size could address the weakness, but the added computational burden makes HAWK less competitive against other post-quantum signature options such as ML-DSA and FN-DSA.
• The AES result is more limited. Mythos also helped improve a meet-in-the-middle attack on a weakened AES challenge instance, reducing the number of required plaintext inputs from about 2^105 to 2^89. That remains far outside practical reach and does not apply directly to full, specification-compliant AES.

Why it matters

The headline should not be read as “AI has broken modern encryption.” The source material is careful about the caveats: the systems tested were either weakened challenge instances or candidates still under review; today’s deployed cryptosystems are not compromised by these results; and the underlying mathematical primitives remain considered safe for now.

The more important development is procedural. Mythos did not invent entirely new mathematics from nothing. Instead, it reviewed literature, reasoned across existing techniques, ran computational experiments, and assembled a more effective attack path. That is precisely the kind of work where advanced AI agents may become useful: cryptographic research often depends on connecting scattered methods and verifying whether a promising construction actually works.

For NIST’s post-quantum process, HAWK’s withdrawal is less a failure than a sign that adversarial review is doing its job. It is far better to identify weaknesses before standardization and deployment than after widespread adoption. For AI research, the episode is more ambiguous. Anthropic’s framing may include marketing, but the result is still significant: if models can generate plausible cryptanalytic advances at scale, the bottleneck may shift from discovery to human verification, triage, and interpretation.

Source: Ars Technica AI