OpenAI outlines security measures for Astra model while Anthropic eases restrictions on Fable

OpenAI announced enhanced security safeguards for its Astra model, a frontier AI system with advanced agentic coding and cybersecurity capabilities. The company will implement isolated testing environments, restricted network and tool access, enhanced monitoring, and sandboxed execution. Additionally, OpenAI introduced universal monitoring of the model's Chain of Thought to detect and interrupt high-risk activities during both training and evaluation. Meanwhile, Anthropic moved in a different direction by relaxing restrictions on its Fable model for biology-related prompts, a shift that reflects competitive pressure as Chinese AI companies develop competitive open-weight alternatives.