Anthropic's Claude AI Finds New Crypto Attacks on HAWK, AES
08 Aug 2026
Anthropic announced on July 28th that its large language model, referred to as Claude Mythos, discovered new cryptanalytic attacks — one against a NIST post-quantum signature candidate, and another against a reduced-round version of AES-128.
What happened
According to Anthropic, Claude Mythos identified an attack on HAWK, a post-quantum signature scheme currently under consideration in NIST's Additional Digital Signatures standardization project. The attack reduces HAWK-512's estimated security level from its 128-bit target to at most 108 bits. A more speculative version of the attack suggests security could fall as low as 81 bits, though this figure carries significant uncertainty.
Separately, Anthropic reported an improved key-recovery attack on a 7-round version of AES-128 — notably not the full-round cipher used in production systems today.
Anthropic also helped create CryptanalysisBench, a benchmark of cryptanalysis tasks spanning full- and reduced-round AES, ChaCha, and algorithms from NIST competitions. In its paper on the AES attack, Anthropic said it intends to further explore LLM-assisted cryptography, particularly in cases where resulting attacks are computationally intractable to actually implement.
What's still unclear
The report leaves several open questions. It's unclear what distinguishes Claude Mythos from publicly available Claude models, how CryptanalysisBench was validated, or who else uses or contributes to it. Whether the 7-round AES-128 attack is practically executable or only a theoretical improvement remains unaddressed, as does whether the HAWK attack is computationally feasible enough to affect its standing in NIST's ongoing standardization process. The speculative 81-bit estimate for HAWK-512 may rely on assumptions that haven't been independently verified.
Why founders should care
For founders building on cryptographic infrastructure — particularly post-quantum signature schemes or systems using reduced-round ciphers — this development is likely worth monitoring rather than immediately acting on:
- Post-quantum standardization risk: If HAWK's reduced security margin holds up under further scrutiny, it could affect the scheme's viability as a NIST standard candidate. Startups evaluating or building on post-quantum crypto may want to track NIST's response before committing engineering resources.
- AI as a cryptanalysis accelerant: The use of an LLM to find these attacks suggests AI-assisted cryptanalysis could increasingly surface weaknesses in cryptographic algorithms — potentially before or alongside traditional research methods. This may signal a growing need for security teams to incorporate AI-driven auditing into their threat models.
- Tooling opportunity: The creation of CryptanalysisBench points to rising interest in benchmarking LLMs on security-relevant tasks. This could open opportunities for security-focused startups building evaluation or auditing tools around AI-assisted cryptanalysis.
- Limited near-term risk to production systems: Because the AES attack applies only to a reduced 7-round version — not the full cipher — and Anthropic has expressed interest in attacks that remain computationally intractable to implement, the immediate practical risk to systems using standard AES-128 appears limited, though this is not certain given the pace of related research.
Founders relying on reduced-round cryptographic implementations, or building products around post-quantum signature schemes, should treat this as an early signal to watch NIST's standardization process and further LLM-assisted cryptanalysis research rather than an immediate call to change existing systems.