AINEWS
HAWK is a proposed, not yet standardized post quantum signature scheme. The attack does not transfer to Falcon, to ML DSA (the standard NIST has already selected), or to other signatures on the standards track.
Anthropic released two new cryptanalysis results on July 28, both attributed to Claude Mythos, the company's still-unreleased research model. The headline result is a new key-recovery algorithm against HAWK, a proposed post-quantum signature scheme that had reached a meaningful stage in the NIST signature-evaluation process. The result does not break anything that ships today.
HAWK is a candidate signing scheme, not a deployed one. It is built on a relatively new algebraic assumption called the module Lattice Isomorphism Problem (module-LIP), which is related to but distinct from the hard problem underlying Falcon, the lattice scheme that is actually being standardized as part of NIST's post-quantum suite. HAWK's pitch to cryptographers has been efficiency, and it had advanced far enough in the evaluation pipeline that a successful attack would have a real effect on the standards conversation, not a hypothetical one.
The attack is still exponential time, so it does not place HAWK in the "broken" bucket. It does, according to Matthew Green's analysis on the Cryptography Engineering blog, roughly halve the number of bits of security the scheme delivers, and the team produced code that runs in a few hours of wall-clock time against a weakened challenge instance provided by the HAWK authors. In practice, the security gap is fixable in principle by doubling the key size, but doubling the key size eats directly into the efficiency advantage HAWK was offering in the first place. The result is a real cryptanalytic advance against a scheme that was actually in the running, not a fracture in deployed cryptography.
The second result, an improved attack against reduced-round AES, is a smaller beat. AES has been under continuous study for two decades, the academic literature on reduced-round variants is dense, and the gap between "reduced-round" and the full cipher is large. The two papers together read as a capability demonstration: HAWK is the load-bearing piece, AES is evidence the same research pipeline generalizes beyond a single scheme.
HAWK is especially tractable for an AI-assisted pipeline because the module-LIP structure exposes patterns a code-generating model can search and exploit in ways earlier tooling could not. Green argues the result is especially ripe for AI help, a read that matches how Anthropic has separately described the work. The same source that flagged the result as ripe for AI help is the one insisting the attack does not transfer to Falcon, because Falcon relies on a different hard problem (NTRU, a lattice problem in the same family but with a different algebraic structure), and the module-LIP feature HAWK uses is not part of Falcon's design.
The Hacker News thread on the post is where readers are working through the cost of a forced key-size doubling and what it does to HAWK's value proposition, with the active arguments staying inside the AI-tractable-algebra frame rather than treating the result as a break of deployed crypto. That matches Green's own read: a real result against a non-deployed scheme, with the standards-track cousins unaffected and the deployed ecosystem intact.
Two things to watch. First, whether the HAWK authors respond with a parameter change and a fresh security claim, or let the scheme drift out of the running. Second, whether NIST's signature-selection process absorbs this as a footnote on a candidate that was almost-but-not-quite in, or as a marker for how aggressively to evaluate the next generation of lattice proposals. Either way, the pattern worth tracking is the narrower one: an AI-tractable algebraic structure produced a security-bit halving that is fixable in theory but economically meaningful in practice.