Quoting Anthropic Frontier Red Team

T2Covered by 1 source · first reported by Simon Willison's Weblog at 29 Sept, 22:20 UTC

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 so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful…