Quoting Anthropic Frontier Red Team

benchmarksSimon WillisonSep 29, 2026

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...

Impact: HighConfidence: MediumNovelty: Material
benchmarkssimon-willisonresearch

See the whole story

Read the original on Simon Willison

Sens.ai aggregates and assesses published reporting. The assessment above is machine generated; the original source is authoritative.