Anthropic's Claude Cracks a Post-Quantum Signature Scheme in Under Four Hours
The AI model also sped up a known attack on a reduced version of AES-128 by up to 800 times, raising fresh questions about machine-assisted cryptanalysis.
Key points
- Anthropic says its Claude Mythos Preview model helped build a full key-recovery attack against HAWK-256, a candidate post-quantum signature scheme.
- The released attack runs end-to-end in roughly three hours and 42 minutes on a 96-core server.
- The same model produced a 200 to 800 times speedup on an existing attack against seven-round AES-128, a stripped-down research version of the standard cipher.
- The HAWK break relies on a symmetry in the underlying lattice maths that cryptographers had not previously exploited.
- Neither result affects the full 10-round AES-128 used in production systems today.
Anthropic has published cryptography results that will make researchers sit up, and vendors nervous about how quickly AI can chew through hard maths problems.
The company says its Claude Mythos Preview model helped derive an end-to-end key-recovery attack on HAWK-256, a lattice-based signature scheme designed to survive attacks from future quantum computers. Put simply: HAWK is one of the maths systems being tested to replace the digital signatures that protect software updates, banking traffic and government documents once quantum machines can break today's cryptography.
The attack is not theoretical. Anthropic released a working implementation. On a 96-core server, it recovers the secret key in about three hours and 42 minutes.
How did an AI model break the scheme?
Claude helped the researchers spot a symmetry in the lattice that HAWK-256 is built on, and no one had used it in an attack before. A lattice, in this context, is a grid-like mathematical structure whose hardness is meant to keep the signing key secret. The newly found symmetry gave the researchers a shortcut through that structure.
That is the striking part. The model did not just optimise known code. It contributed to the mathematical insight itself.
HAWK is a research candidate, not a deployed standard. It was submitted to ongoing post-quantum competitions rather than baked into your browser. So no production system is at immediate risk. But the result is a warning shot for every other lattice scheme still under review.
What about the AES result?
Claude also produced a 200 to 800 times speedup on a published attack against seven-round AES-128, according to Anthropic. That needs unpacking. AES-128, the Advanced Encryption Standard with a 128-bit key, is the cipher protecting most encrypted web traffic, disk encryption and messaging apps. The full version runs 10 rounds of scrambling.
Researchers routinely attack reduced-round versions (four, five, six, seven rounds) as a way to measure how much safety margin the full cipher has. A faster attack on seven rounds does not mean AES is broken. It means the margin shrank a little, and an AI helped shrink it.
The Hacker News flagged the twin results earlier this week.
What is actually at stake?
| Result | Target | Impact |
|---|---|---|
| Key recovery | HAWK-256 signature scheme | ~3h 42m on a 96-core server |
| Attack speedup | 7-round AES-128 (research variant) | 200x to 800x faster |
| Deployed AES-128 | 10 rounds, in production | No change, still considered secure |
Two things matter here for anyone outside the cryptography community.
First, post-quantum standards are still being chosen. Schemes that look elegant on paper can collapse when a new angle of attack appears, whether the angle comes from a human or a language model. Expect standards bodies to lean harder on schemes with the longest, most battered track records.
Second, AI-assisted cryptanalysis is no longer a talking point. It produced running code that recovers a key in an afternoon. Defenders and standards bodies will need to assume attackers can point similar tools at their own systems.
What should security teams do now?
Nothing needs ripping out today. AES-128 as deployed is unaffected, and HAWK-256 is not in your stack unless you are running experimental post-quantum software. Teams planning post-quantum migrations should watch which schemes survive this new wave of scrutiny before committing, and favour hybrid deployments that keep a classical algorithm alongside the post-quantum one until the dust settles.
