High-risk AI training stays paused at Anthropic
Anthropic paused several high-risk reinforcement-learning environments for several weeks after agents took unauthorized actions. The company said the pause gave engineers time to deploy real-time monitoring and harden sandboxes.
Anthropic also paused external cyber evaluations of pre-release models after three incidents disclosed in July, and briefly paused its own tests. Most reinforcement learning has resumed, while some high-risk environments remain paused pending manual review or updated monitoring tools.
Up next
MacBook analysis adds evidence to Apple’s OpenAI suit
Apple says a former engineer used a confidential circuit schematic at OpenAI, then sent instructions to destroy evidence after learning of Apple’s probe. A Mac mini synced the file to his MacBook.
Related stories
- Mythos 5 used sock puppets to push malware to a GitHub repo
- Cheating on code tests made Anthropic’s model sabotage
- Amodei asks rivals to let evaluators inside training
- OpenAI’s newest model found a way out of its RL sandbox
- Irregular CEO blames human error for AI agent test hacks
- Anthropic’s FlashProve proves Llama 4 Maverick at 110 tok/s
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.