Anthropic Reports AI-Assisted Attacks on HAWK-256 and Seven-Round AES-128

Anthropic says its Claude Mythos Preview research system helped develop two cryptanalytic results: an end-to-end key-recovery attack against the HAWK-256 challenge parameter and a substantially faster attack against a seven-round version of AES-128.

Neither result is considered a practical threat to deployed systems. The HAWK work targets a smaller parameter set created as a cryptanalysis challenge, rather than HAWK-512 or HAWK-1024, the security-level parameters under consideration in a National Institute of Standards and Technology post-quantum signature process. The AES technique covers only seven of AES-128’s 10 rounds and requires roughly 2105 chosen plaintexts.

Exploiting a lattice symmetry

The HAWK attack builds on earlier research showing that a hidden symmetry in the scheme’s lattice could reduce the key-recovery problem to a shorter-vector search in a lower-dimensional structure. Anthropic said Claude identified an additional automorphism that made the approach operational.

The resulting method constructs a specialized lattice from a public key, applies reduction and sieving techniques, and derives alternative signing material. The released implementation reportedly takes an estimated three hours and 42 minutes on a 96-core server under the HAWK-256 conditions tested. It does not reproduce the original secret-key seed, but generates an equivalent key that can sign messages accepted by the reference implementation.

Anthropic estimates that the work factor for HAWK-256 falls from about 264 to 238. It also reported lower theoretical estimates for the larger parameter sets, while emphasizing that those attacks remain computationally infeasible. The public code supports only HAWK-256 and includes demonstration keys and tools for testing newly generated keys.

Reducing an AES attack cost

For seven-round AES-128, Claude reportedly found an invariant that eliminates a 256-value guessing stage from an existing meet-in-the-middle strategy. Anthropic calls the construction the “Möbius Bridge” and estimates a speedup of 200 to 800 times, depending on the measurement basis.

The published artifacts validate individual components and perform full recovery on a smaller AES-like cipher. They do not carry out a complete attack on real seven-round AES-128; the projected result remains impractical because of its data requirements and scale.

Anthropic said the AI performed much of the exploratory work, while researchers supplied direction, computing resources and extensive verification. The company reported approximately $100,000 in API costs and several hundred hours of human review. It said the findings do not require changes to production software.