Claude uncovers flaws in HAWK and reduced AES, but no cause for alarm

On July 28, 2026, Anthropic’s Frontier Red Team published research detailing cryptographic weaknesses identified using an internal preview of Mythos, a model not yet publicly available.

Post-quantum hawk takes a brief flight

The HAWK signature scheme was, until recently, a candidate in the NIST’s post-quantum cryptography standardization process. The goal was to identify algorithms to complement the four winners selected in 2022. HAWK had successfully passed two rounds of expert review over two years.

Breaking post-quantum candidates is not unprecedented. SIKE was famously cracked in under an hour on a standard PC, and Rainbow, another finalist, saw its secret key recovered after an average of 53 hours. In the case of HAWK, Mythos leveraged the best-known attack against the scheme, refining it over approximately 60 hours of work. This reduced HAWK’s effective security by half in terms of key size.

The technical details are complex, but HAWK’s security relies on the difficulty of the Modular Lattice Isomorphism Problem. Mythos discovered a geometric symmetry within this problem, enabling a reduction in the search for the private key to a short vector problem in a lattice with halved dimensions. For the HAWK-256 variant, this translated to a complexity drop from 2⁶⁴ to 2³⁸ operations. In practice, the secret key was extracted in just 3 hours and 42 minutes on a standard 96-core server.

This halving of HAWK’s effective security undermines its practical value. To restore an acceptable level of protection, designers would need to double key and signature sizes, compromising the lightweight computation that was HAWK’s primary advantage. As a result, the HAWK team officially withdrew the algorithm from the NIST selection process on July 29, rendering it obsolete in its current form.

What about AES?

The other focus of Anthropic’s research is AES, a widely deployed encryption standard used in banking, TLS, and supported by hardware acceleration in Intel and AMD processors. To test AES’s robustness, researchers often use reduced-round versions of the algorithm. Mythos found a way to break one such variant by eliminating a key assumption for attackers, improving the speed of the best-known attacks by a factor of 200 to 800.

This breakthrough applies strictly to a modified, simplified version of AES-128, reduced to just 7 rounds instead of the standard 10. The innovation lies in a technique called the Möbius Bridge, an invariant fingerprinting method integrated into a meet-in-the-middle attack. This approach eliminates a previously exhaustive search step requiring 256 iterations, drastically accelerating the attack’s execution.

Anthropic emphasizes that this vulnerability remains confined to academic research and is impractical in real-world scenarios. Even against the weakened AES variant, the attack would be challenging, and the additional three rounds in the full AES standard make it even more infeasible.

France’s National Agency for the Security of Information Systems (ANSSI) recommends a minimum symmetric key size of 128 bits for standard security and 192 bits for post-quantum resistance. While increasing key sizes may slightly impact performance, the overhead is minimal, as seen with AES-192 compared to AES-128.

Food for thought

While these findings have no immediate impact on current systems—HAWK was not in use, and AES remains secure—they raise important considerations. The loss of HAWK represents the failure of a potential post-quantum algorithm, though Anthropic notes that other candidates remain unaffected. However, the episode highlights the increasing role of AI in cryptanalysis.

Discovering HAWK’s weakness required approximately 60 hours of autonomous computation, costing around $100,000 in API fees if using Anthropic’s commercial tools. Mythos initially refused to execute the task on the reduced AES variant, claiming no improvements were possible on such a well-established algorithm. After researchers adjusted the model’s prompting framework (a process Anthropic calls “scaffolding”), the AI rewrote its own execution framework, generating about a billion tokens over several days to devise a new attack method. Again, this would have cost $100,000 in standard conditions.

The results were not published immediately upon generation by the LLM. Anthropic states that two researchers reviewed the findings for a month, verifying the mathematical rigor of the demonstrations and testing the code generated by the AI agent. This process aimed to eliminate potential hallucinations and validate every equation, algebraic transformation, or complexity reduction proposed by the model.

The bottleneck shifts again

Even if these discoveries have no concrete impact on current computing (broadly speaking), the cybersecurity community is under immense pressure. LLMs are uncovering so many vulnerabilities that human processes—triage, verification, and remediation—are struggling to keep up. The surge in patches for products like Firefox, Windows, and Apple systems is a testament to this challenge.

Anthropic’s research underscores a shifting bottleneck: the problem is no longer finding flaws but validating the results produced autonomously. The time spent conceptualizing the idea was about a week, followed by a month of verification by two researchers. This raises the question: could a human researcher have reached the same conclusions through relentless persistence?

This is not an isolated incident. Anthropic has hinted at preliminary findings on other algorithms, signaling that more publications may follow.