Anthropic’s Frontier Red Team finds GLM-5.3 and Claude Mythos Preview can hijack code
Anthropic’s Frontier Red Team tested two models—GLM-5.3 and Claude Mythos Preview—on 100 tasks from its internal Binary Exploitation benchmark. GLM-5.3 succeeded in 4% of trials, while Claude Mythos Preview achieved 6%. Both models demonstrated advanced cyber capabilities, unlike earlier versions like Claude Opus 4.6 and GLM-5.2, which failed entirely.
Key points
- GLM-5.3 hijacked code in 4% of 100 trials, per Anthropic’s Frontier Red Team
- Claude Mythos Preview succeeded in 6% of the same tests, per the same report
- Earlier models like Claude Opus 4.6 and GLM-5.2 failed all trials, per the team
The findings suggest a significant shift in model behavior, raising concerns about the spread of dangerous capabilities. Anthropic’s report does not specify which companies developed GLM-5.3 or Claude Mythos Preview, nor does it assess broader real-world risks beyond controlled tests.
Model page: GLM-5.3 →
The story so far
2 episodes →- Anthropic’s Frontier Red Team finds GLM-5.3 and Claude Mythos Preview can hijack codethis story
Quoting Anthropic Frontier Red Team
Simon Willison · 29 September 2026Loading the full article…
This text was published by Simon Willison and written by Simon Willison. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Swift-1.5-Qwen3.8-27b-oQ8e-mtp achieves 34.8 tokens per second on Apple M5 Max · 1 src
- Zhipu automates infrastructure with GLM-5.3’s outer RSI loop in under two weeks · 3 src
- Ohio State and Google launch AI research hub with DeepMind models and Gemini access · 1 src
- General Intuition raises $220M at $6.2B valuation for AI agents trained on gameplay footage · 1 src
- Anthropic’s AI finds a CRISPR-like gene-editing system in phages · 1 src
Comments
via GitHub Discussions