Anthropic Frontier Red Team reports first successful control-flow hijacks by GLM-5.3 and Claude Mythos Preview
Evaluation on an internal binary exploitation benchmark finds early models failed all tasks, while newer releases achieved non-zero success rates.
2 sources · cross-referenced
- Anthropic’s Frontier Red Team evaluated GLM-5.3 and Claude Mythos Preview on 100 randomly selected binary exploitation tasks.
- GLM-5.3 achieved full control-flow hijacks in 4% of trials; Claude Mythos Preview in 6%.
- Earlier models, including Claude Opus 4.6 and GLM-5.2, succeeded in none of the trials.
Anthropic’s Frontier Red Team describes an evaluation on 100 randomly selected tasks from an internal binary exploitation benchmark. The team reports that GLM-5.3 achieved full control-flow hijacks in 4% of trials, while Claude Mythos Preview did so in 6%.
Although GLM-5.3 underperformed relative to Claude Mythos Preview on this benchmark, Anthropic characterizes the results as crossing a meaningful threshold: earlier models such as Claude Opus 4.6 and GLM-5.2 did not succeed in any of the trials.
The Frontier Red Team’s findings are presented as part of a broader research post titled 'GLM-5.3 and the spread of advanced cyber capabilities.'
- Oct 13, 2026 · Google DeepMind — Blog
Google DeepMind releases Gemini 4 Argon with domain-leading performance and multi-domain benchmarks
Trust79 - Oct 9, 2026 · Hugging Face
Hugging Face publishes Falcon ASR, a 1.6B-parameter Arabic speech recognition model with Emirati focus
Trust84 - Oct 8, 2026 · OpenAI — News
OpenAI disrupts two AI-enabled influence operations using its models
Trust79