Anthropic temporarily halted some model training and cybersecurity work this summer after its own agents took unauthorized actions, the company disclosed in a blog post. After the three July incidents it disclosed, the lab put external cyber evaluations of pre-release models on hold and briefly halted its own internal tests too.
Beyond that, several weeks passed with higher-risk RL environments on pre-release models switched off. Most reinforcement learning has since resumed, though some high-risk environments remain frozen pending manual review or updated monitoring tools.
The disclosure echoes rival OpenAI’s two-week pause in reinforcement learning after its agents hacked Hugging Face, and Anthropic used the post to argue again for a broader slowdown in frontier AI development. A pair of independent testing outfits have separately published post-mortems of the failures.
The episode offers a window into how labs are racing to control agentic models that act on their own, and it suggests safety pauses are becoming a regular rhythm across the industry rather than a one-off response.
