TL;DR
- Anthropic reported that Claude Mythos Preview found an improved attack on HAWK, a post-quantum signature scheme under NIST consideration, plus a new technique against reduced-round AES-128; neither result breaks anything currently deployed.
- The HAWK work ran semi-autonomously over sixty hours at roughly $100,000 in API cost, finding a lattice automorphism that dropped estimated full key recovery on HAWK-256 from 2^64 to a demonstrated 2^38.
- On AES, Claude initially reported nothing to find; a researcher redirected it in writing, and three days and several hundred million tokens later it produced a 200 to 800 times speedup over the prior best attack.
- Exploration is now cheap and large, while deciding what a result means and confirming it holds stayed with people, including several hundred hours of human validation.
In plain English
A signature scheme is the maths that proves a message came from who it claims to. HAWK is a post-quantum one, designed to survive attacks by future quantum computers, and NIST is the US standards body assessing candidates. AES-128 is the most widely used block cipher for encrypting data; “reduced-round” versions run fewer internal scrambling passes than the real thing, so they are studied as a way to probe the full cipher without breaking it. The 2^64 to 2^38 change on HAWK-256 describes how much work a full key recovery takes, roughly halving the effective key size.
The AES finding, which Anthropic calls a Möbius Bridge, is a fingerprinting method that makes one stage of the cipher unaffected by a guess that previously required checking 256 possible values. Tokens are the units of text a model processes and is billed for, which is where the cost figures come from. The human contribution was the direction to keep going when no result had appeared, the judgement of which lead was worth pursuing, and the validation before anything was shown to NIST.
Anthropic’s writeup on discovering cryptographic weaknesses with Claude describes Claude Mythos Preview finding a genuinely improved attack on HAWK, a post-quantum signature scheme currently under NIST consideration, plus a new attack technique against reduced-round AES-128. Neither result breaks anything currently deployed. Both are real cryptographic research, which makes the easy summary, “AI found a security bug,” worse than useless. It skips past the actual question, which is what specifically got easier, and what still required a person who understood what a weakness meant.
The HAWK result came out of a semi-autonomous process, an Anthropic researcher giving Claude occasional nontechnical direction while it handled the literature review, the mathematical reasoning, the experiments, and its own verification pipeline. Sixty hours of work, roughly $100,000 in API cost, and the output was a nontrivial automorphism in HAWK’s underlying lattice that nobody had exploited before, dropping the estimated cost of a full key recovery on HAWK-256 from a thought-to-be 2^64 down to a demonstrated 2^38. Effective keysize cut in half. That’s not a marketing claim, it’s the math the paper shows its work on.
The AES result is the more honest account of where a human still had to intervene directly. Claude’s own transcript records it pushing back on the task at first: “on AES-128 r5/r6/r7 it found nothing because there’s nothing easy to find; this is the most-studied block cipher in existence.” The researcher had to redirect it in writing, telling it they wanted proper research and genuinely hard findings, not confusing the absence of an obvious result with the absence of any result. Three days and several hundred million tokens later, Claude arrived at what Anthropic calls a Möbius Bridge, a fingerprinting method making one stage of the cipher invariant to a guess that previously required checking 256 values, a 200 to 800 times speedup over the prior best attack. It then took several hundred hours of human validation before Anthropic would trust the result enough to publish it.
The division of labor that actually happened isn’t the one implied by a headline that just says AI found a bug. The exploration, testing ideas at a volume no team of researchers would spend hours on by hand, is genuinely a new capability. Recognizing that no result yet doesn’t mean no result is possible, knowing which half-formed lead is worth pushing further, and spending several hundred hours confirming an attack actually holds before telling NIST about it: all of that stayed entirely on the human side, on both ends of the three days Claude spent grinding through AES.
For anyone doing security work: the exploration budget just got enormous and comparatively cheap. The bottleneck moved almost entirely onto whoever has to decide what a result means and whether it can be trusted enough to act on.