Anthropic says Claude improved attacks on HAWK and reduced-round AES

Why it matters

Claude has moved from finding implementation bugs to proposing attacks on cryptographic designs. Founders, security teams and infrastructure operators need expert validation, responsible-disclosure procedures and model-access controls before similar methods target deployed systems.

Anthropic, co-founded by Dario Amodei and Daniela Amodei, said on July 28 that its controlled-access Claude Mythos Preview model helped researchers improve the best-known attack on HAWK, a post-quantum signature candidate that advanced to the third round of a US government standardization process in May. Anthropic says the work effectively cut HAWK's key strength in half.

Editor's note: This article covers Anthropic's disclosed HAWK and reduced-round AES findings and their release process; a separate RuntimeWire story covers CryptanalysisBench as an evaluation dataset.

The disclosure gives the Amodeis a concrete test of the bargain behind Anthropic's security strategy. The same model capabilities that can find serious software flaws can also probe the mathematics protecting digital communications. For founders and security leaders, the research raises practical questions about who can use cyber-capable models, how their findings should be verified and when attacks should be released.

Daniela Amodei has described Anthropic's origin around building AI systems safely "from day one." Dario Amodei, a Princeton-trained biophysicist who previously led research at OpenAI and worked at Google Brain, has built Anthropic around the view that capability gains and risk controls have to be developed together. Cryptanalysis makes that tension unusually plain: a successful attack can help cryptographers reject a weak design before deployment, while the same ability could become dangerous against systems already in use.

Anthropic's research post describes two principal results. Mythos improved the best-known attack on HAWK after about 60 hours of work, Anthropic says. A separate run found a method that made an existing attack against reduced-round AES between 200 and 800 times faster. Anthropic estimates that each result consumed roughly $100,000 in API usage.

Neither finding requires operators to replace software or rotate keys, according to Anthropic. HAWK has not been deployed as a standard, and the AES work targets a weakened research variant. Full AES remains unaffected.

Claude found a symmetry in HAWK

HAWK is a lattice-based digital signature scheme designed for post-quantum security.

NIST advanced HAWK alongside eight other candidates on May 14 after two rounds of review. HAWK remains in public review as a candidate and is not a deployed NIST standard.

Anthropic says Mythos found an exploitable symmetry, known as a nontrivial automorphism, in the lattice underlying HAWK. Earlier research had shown that efficiently finding such an automorphism would permit an attack. Mythos identified one that applied to HAWK, leading Anthropic to describe the result as effectively cutting the scheme's key strength in half.

The finding arrived during the public review process intended to expose weaknesses before browsers, identity systems, hardware vendors and cloud services build around a standard.

Anthropic said it shared the attack with HAWK's authors in June, consulted academics and provided advance copies to US government and industry partners before publication.

The AES result tests a heavily scrutinized cipher

The second result concerns a reduced-round version of AES, while production systems use the complete cipher. Researchers study reduced-round ciphers to examine the security margin protecting the full design and test techniques that could inform later work.

Anthropic says Mythos eliminated one guess required by an existing attack, producing a calculated speed improvement of 200 to 800 times over previous methods. Anthropic is explicit that the result applies only to the reduced-round research variant and does not break full AES.

According to Anthropic's disclosure, Mythos completed most of the HAWK and AES work autonomously, with limited human intervention. Anthropic relied on academic consultation and responsible disclosure procedures to validate and release the findings. That process puts pressure on the slower parts of AI-assisted security research: expert review, novelty checks and decisions about when and how to publish a credible attack.

RuntimeWire reported earlier in July that Mythos-driven vulnerability disclosures were already exposing a patching bottleneck. Algorithm research poses a different verification problem because a mathematical argument may demand specialist review beyond a crash report or proof-of-concept exploit.

Anthropic ties the work to controlled access

When Anthropic introduced Mythos Preview on April 7, Anthropic said the model could autonomously find and exploit vulnerabilities in software it tested, including several major cryptographic libraries. Those claims concerned implementation mistakes in software. The HAWK and AES work extends the model's reported capabilities to weaknesses in cryptographic designs.

Anthropic initially positioned Mythos Preview within Project Glasswing, its initiative for securing critical software. The cryptography disclosure shows why Anthropic is emphasizing controlled access for cyber models: the research capability can produce attack methods that require specialist review and a managed release process.

The Amodeis' security thesis now has an algorithm-level test. The immediate operational impact is limited, but founders and security leaders must decide how to validate, restrict and responsibly disclose model-generated attacks before similar methods reach cryptography carrying production traffic. Models that identify weak designs during standardization could help vendors avoid costly deployments. Specialist review still determines whether a proposed attack is credible.