Anthropic's Claude Mythos Preview has done something that would have taken a team of cryptographers months, if not years: it derived a full key-recovery attack against HAWK-256, a post-quantum digital signature scheme, and separately found a speedup of between 200 and 800 times for an attack on seven-round AES-128. Anthropic has published the implementation. On a 96-core server, the HAWK attack completes end-to-end in under four hours.

To understand why this matters, a quick detour. The world is in the middle of a multi-year transition away from classical cryptographic standards — RSA, elliptic curve, and their kin — toward post-quantum algorithms designed to resist attacks from future quantum computers. NIST has been running a years-long competition to standardise these new schemes. HAWK-256 is one of the candidates still in the running. It's not yet a deployed standard, but it's exactly the kind of algorithm that organisations are beginning to evaluate for adoption.

The attack Claude found exploits a symmetry in the mathematical lattice structure underlying HAWK's signature scheme — a property that, according to Anthropic, had not been previously identified in published research. That's not a subtle incremental improvement on a known weakness; it's a novel finding in a field where novel findings usually come from dedicated academic research groups working for months.

The AES result is more nuanced. AES-128 at seven rounds is not full AES — the deployed standard uses ten rounds — so this doesn't mean your encrypted hard drive is suddenly vulnerable. But cryptanalytic speedup on reduced-round AES is still a meaningful signal: it suggests AI-assisted cryptanalysis is beginning to close gaps that have historically been intractable.

Taken together, these results are a proof of concept for something the security community has been nervously anticipating: AI as a capable cryptanalyst, not just a code generator or vulnerability scanner. Prior AI-assisted security work — finding bugs in software, accelerating exploit development — operated in a domain where human intuition already had good footholds. Pure mathematics, especially lattice cryptography, has been more resistant to that kind of acceleration.

This also lands awkwardly for Anthropic. The company has positioned itself as the safety-first AI lab, and the FBI has already flagged Mythos as a law enforcement challenge given its capabilities in sensitive domains. Publishing the HAWK attack implementation is arguably the responsible disclosure path — the scheme's designers can now respond, and the wider community can assess the damage. But it's also a demonstration that the most capable AI systems are now capable of original mathematical research with direct security implications.

What to watch: How HAWK-256's designers respond, and whether NIST's post-quantum evaluation process accelerates scrutiny of other lattice-based candidates. If Claude can do this with HAWK, the obvious question is which other candidate schemes have unexplored symmetries waiting to be found.

For Australian organisations in the early stages of post-quantum roadmapping — a process the ASD has been nudging critical infrastructure operators toward — this is a useful reminder that "candidate algorithm" does not mean "safe to deploy."