Anthropic's Claude AI Finds Faster Attack on HAWK-256 and Seven-Round AES-128

A recent breakthrough by Anthropic’s Claude AI has revealed a faster attack on the lattice-based signature scheme, HAWK-256. The research also uncovered an improved method for attacking seven-round AES-128, a widely used encryption standard. According to the findings, the new attacks are significantly faster than previous methods but still remain impractical in real-world scenarios.

The HAWK-256 attack exploits a previously unused symmetry in the lattice behind the signature scheme. This discovery was made possible by Claude AI’s ability to analyze and manipulate complex mathematical structures. The resulting attack is an end-to-end key-recovery method that can recover short vectors from the public key, allowing for the reconstruction of a secret basis.

The researchers used Anthropic’s Mythos Preview tool to derive the HAWK-256 attack. This tool largely conducted the research itself, with human input limited to project direction and occasional guidance. The entire process took approximately 60 hours in a multi-agent environment, with an estimated API cost of around $100,000.

The new attack on seven-round AES-128 targets one of the ten rounds of the encryption standard. By removing a 256-way guessing step from existing meet-in-the-middle attacks, Claude AI’s method achieves a speedup of between 200 to 800 times faster than previous methods. However, this improvement still requires an impractical number of chosen plaintexts and remains infeasible at realistic scales.

The researchers emphasized that neither the HAWK-256 nor seven-round AES-128 attack affects production systems or poses any immediate threat to security. The company stated that no changes are needed for existing software as a result of these findings. Additionally, both larger parameters (HAWK-512 and HAWK-1024) remain impractical to attack.

HAWK is the only lattice-based scheme among nine candidates advanced by NIST in its post-quantum standardization process. Its security parameter sets are HAWK-512 and HAWK-1024, with HAWK-256 serving as a challenge parameter for cryptanalytic purposes. The attack on HAWK-256 targets this smaller parameter set rather than the larger ones.

The researchers used Claude AI to find an additional automorphism in the lattice that enables the exploitation of previously discovered symmetry. This discovery was made possible by the tool’s ability to analyze and manipulate complex mathematical structures. By constructing a τ-cocycle lattice from the public key, the attack can recover short vectors before reconstructing a secret basis.

The HAWK-256 recovery process involves verifying the recovered key through signing a message with the NIST reference implementation. However, this does not reveal the original 96-byte secret-key seed but instead produces a decoded key containing functionally equivalent signing material. The company’s released code supports only HAWK-256 and rejects non-HAWK-256 inputs.

The AES result is narrower in scope than its name might suggest. While it targets seven rounds of AES-128, the attack remains impractical due to the requirement for an impractically large number of chosen plaintexts. The researchers used Claude AI’s Möbius Bridge invariant fingerprint to remove a 256-way enumeration step from existing meet-in-the-middle attacks.

The development and verification process took several hundred hours, with two researchers spending nearly a month to reach confidence in the correctness of the method. Anthropic estimated that the expected HAWK-256 key-recovery work factor falls from 264 to 238, while gate-count estimates for larger parameters decreased by around 42.

The company’s disclosures follow the release of CryptanalysisBench, a benchmark developed by researchers from ETH Zurich and other institutions. This benchmark evaluated Mythos 5, which Anthropic describes as an update to its earlier tool, Mythos Preview. The HAWK and AES disclosure specifically names Mythos Preview as the tool used in these attacks.

The public thread on NIST’s forum announcement contained no replies when checked, and there is currently no independent reproduction of Anthropic’s HAWK-256 recovery available for review.