Story
simon_willison · Sep 29, 2026 · news
Source brief
Quoting Anthropic Frontier Red Team
simonwillison.netSep 29, 2026
original source linked
In brief
We evaluate several models on 100 tasks from the [internal Binary Exploitation benchmark] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did s...
Feed lens
eval