What's this about?
Anthropic reported that its "Claude Mythos Preview" model helped its researchers find genuine weaknesses in cryptographic algorithms — the math that keeps data private. It slightly improved a known attack on a reduced version of AES (the standard encryption used everywhere), and, more interestingly, cut in half the effective key strength of HAWK, an algorithm being considered for a future encryption standard. One attack was discovered by the AI running largely on its own.
What exactly is "finding a cryptographic weakness"?
Encryption algorithms scramble data so only someone with the key can read it. Cryptographers "attack" them not by breaking real systems but by finding math shortcuts that make cracking easier than it should be — e.g. reducing how much work, or how long a key, is needed. "7-round AES-128" is a deliberately weakened lab version (real AES uses 10 rounds), used to measure how close attacks are getting — so improving an attack on it doesn't mean your bank is exposed. Halving HAWK's effective key length is a bigger deal because HAWK is a live candidate for standardization, not a toy.
What's the community saying?
- Reassurance first: a top comment spelled out "AES is NOT broken — this is 7 of 10 rounds," heading off panic; the HAWK result is the one worth watching.
- The cost detail everyone quoted: each result took ~$100k in API tokens over about a week — a striking data point on what autonomous research actually costs.
- A philosophical take: applying huge effort to a crypto scheme without breaking it "hardens" it — failed attacks are evidence of strength, so this kind of AI red-teaming is useful either way.
- Multi-agent dynamics: the writeup noted the key idea came from two AI workers collaborating — one gave up on the idea early, the other found a way to make it work — which readers found a fascinating glimpse of emergent teamwork.