Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-31 · high
It also found flaws in a deliberately weakened AES variant, with persistence, not capability, cited as the real limit.
Why now: Anthropic's account of the exercise became public on July 28, the clearest look yet at how long a model will stay on a problem with minimal supervision.
Claude Mythos spent about 60 hours finding cryptographic weaknesses in the HAWK signature scheme and a deliberately weakened AES variant, per Anthropic researchers.
For teams deciding how much autonomy to give AI systems on research work, this is the number that matters. The model didn't stall on the hard cryptography. It stalled on staying with the problem, and someone had to keep telling it to continue.
Human involvement was mostly encouragement, per the report. Researchers told the model not to give up and to push for publishable quality, not steering the technical work itself.
One caveat: the AES variant Claude found flaws in was deliberately weakened for the exercise, not production AES. That's not evidence real-world AES is at risk. The report doesn't say whether the same persistence holds when nobody checks in at all.
The more interesting read: capability was apparently never the constraint. Someone had to keep the model on task, and that's a much smaller problem to solve than the cryptography itself. Watch whether Anthropic tries the same experiment with zero human check-ins next time.
Anthropic published the account on July 28.
Each link below shares sources, entities, or timing with this story.
Claude Mythos benchmarked against Claude / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Claude Mythos benchmarked against Claude); both cover AES, Anthropic, Claude Mythos, HAWK; overlapping topics (claude, hawk).
Linked by a graph relationship (Claude Mythos benchmarked against Claude); both cover Anthropic, Claude Mythos; overlapping topics (capability, claude, mytho).
Claude Mythos built by Anthropic / Shared entities / Earlier coverage
Linked by a graph relationship (Claude Mythos built by Anthropic); both cover AES, Anthropic, HAWK; earlier AES coverage from 2026-07-30.
Linked by a graph relationship (Claude Mythos built by Anthropic); both cover AES, Anthropic, Claude Mythos; earlier AES coverage from 2026-07-30.
Claude Mythos built by Anthropic / Shared entity: Anthropic / Same source domain / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Claude Mythos built by Anthropic); both cover Anthropic; reported by the same outlet (simonwillison.net).
Linked by a graph relationship (Claude Mythos built by Anthropic); both cover Anthropic; reported by the same outlet (simonwillison.net).
Claude Mythos built by Anthropic / Shared entity: Anthropic / Same source domain / Shared topic / Earlier coverage
Linked by a graph relationship (Claude Mythos built by Anthropic); both cover Anthropic; reported by the same outlet (simonwillison.net).
Claude Mythos benchmarked against Claude / Shared entity: Anthropic / Same source domain / Shared topic / Earlier coverage
Linked by a graph relationship (Claude Mythos benchmarked against Claude); both cover Anthropic; reported by the same outlet (simonwillison.net).