Anthropic red team: GLM-5.3 hits full control-flow hijack in 4% of trials, Claude Mythos in 6%
What happened
Anthropic's Frontier Red Team evaluated several models on 100 randomly selected tasks from an internal binary exploitation benchmark. In a report dated 2026-09-30, GLM-5.3 achieved full control-flow hijacking in 4% of trials, and Claude Mythos Preview in 6%. The report says earlier models, including Claude Opus 4.6 and GLM-5.2, failed all of these tasks, and concludes that a meaningful capability threshold has been crossed.
Written by AI from the coverage · updated 53 minutes ago
Coverage
Follow the reports to see the story from different sides.
- Simon WillisonQuoting Anthropic Frontier Red Team
Anthropic Frontier Red Team 在内部 Binary Exploitation 基准的 100 个随机任务上评测多个模型，GLM-5.3 在 4% 的试验中实现了完整控制流劫持，Claude Mythos Preview 为 6%。报告指出，Claude Opus 4.6 和 GLM-5.2 等更早的模型在这些任务中均未成功，认为一个有意义的能力门槛已被跨过。
Heat over time
Not enough continuous observations to draw a trend yet.