Claude AI Autonomously Discovers Cryptographic Weaknesses That Escaped Expert Review

Researchers at Anthropic reported that Claude Mythos Preview autonomously discovered new cryptographic attacks against the post-quantum signature candidate HAWK and a reduced-round version of AES.

This demonstrates that advanced AI models can now contribute to research that was traditionally the domain of specialized cryptanalysts. While the findings do not compromise currently deployed systems, they raise important questions about how the cryptography community will validate and manage discoveries generated by AI.

Claude AI Discovers Cryptographic Weaknesses

In the first result, Claude identified a previously unexploited lattice symmetry known as a nontrivial automorphism in HAWK’s underlying Lattice Isomorphism Problem.

HAWK is a third-round candidate in NIST’s Additional Digital Signatures process, which seeks cryptographic schemes that are resistant to future quantum computers.

Anthropic stated that the new key-recovery technique reduces HAWK’s effective key strength by roughly half. Specifically, for the HAWK-256 parameter set, the expected cost of a full key-recovery attack decreased from \(2^{642}\) to \(2^{382}\) operations.

Although the attack remains exponential rather than polynomial-time, making it impractical against larger HAWK parameter sets, it does not affect other lattice-based schemes or NIST post-quantum candidates.

However, it may necessitate HAWK doubling its key sizes to restore the intended security margin, potentially undermining the efficiency advantages that initially made it an attractive post-quantum digital signature candidate.

Anthropic reported that a team of AI agents found, developed, and verified the HAWK attack in approximately 60 hours with limited human guidance. The agents utilized Python, Sage, cryptography literature, and a collaborative sandbox environment. Researchers estimated that the work consumed about $100,000 in API usage.

Anthropic coordinated its disclosure with HAWK’s authors in June and subsequently released the findings to the public NIST mailing list along with its research.

Claude also developed an improved meet-in-the-middle attack against a reduced-round version of AES-128. The full AES-128 employs 10 rounds, and the new method does not weaken the production cipher.

The model’s “Möbius Bridge” fingerprinting technique eliminates one of the \(2^{28}\) value guesses needed in previous methods, improving estimated attack performance by between 200 and 800 times under a chosen-plaintext model.

However, this scenario still requires approximately \(2^{105}\) chosen plaintexts, making it impractical outside of academic cryptanalysis.

Research area	Target	AI-discovered improvement	Security impact	Production impact
Post-quantum signatures	HAWK	Identified a nontrivial lattice automorphism enabling faster key recovery	Effective key strength reduced by about half; HAWK-256 estimate drops from 2642^{64}264 to 2382^{38}238	None; HAWK is a NIST candidate and not deployed
Symmetric cryptography	7-round AES-128	“Möbius Bridge” fingerprint removes a 282^828 guess in a meet-in-the-middle attack	Estimated attack acceleration of 200–800×	None; full AES-128 has 10 rounds
Lightweight cipher research	13-round LEA	Preliminary practical key-recovery attack	Claimed recovery with fewer than 2302^{30}230 encrypted plaintexts in under an hour	None; full LEA uses 24 rounds
Block cipher research	6-round Serpent-128	Preliminary full key-recovery attack	Extends prior work against a reduced-round target	None; full Serpent-128 uses 32 rounds
Research infrastructure	CryptanalysisBench	Benchmark developed with academic partners	Enables ongoing evaluation of LLM cryptanalysis capability	Supports research and model testing

The research team noted that Claude initially rejected the AES task as infeasible. However, adjustments to the agent harness and prompting encouraged it to explore novel attack ideas.

Over several days, the system generated hundreds of millions of tokens and ultimately developed the central Möbius Bridge concept. Human researchers then spent hundreds of hours validating the method and preparing the technical paper, indicating that verification, rather than discovery, may become the primary bottleneck.

Anthropic also disclosed preliminary AI-assisted attacks against reduced-round versions of LEA, Serpent-128, Salsa20, Poseidon, and SHA-1 variants.

To help researchers track these capabilities, the company partnered with ETH Zurich, Tel Aviv University, and the University of Haifa on a project called CryptanalysisBench.

These results underscore the value of adversarial review and suggest that AI-assisted cryptanalysis may soon enhance scrutiny of both emerging and legacy cryptographic designs.

ALERT: 20+ government sites delivered malware to businesses and citizens. See full attack research to check your own exposure.