OpenAI has suspended the development of its Astra model after internal evaluations showed significant advancements in agentic coding and cybersecurity capabilities. The company cannot rule out that the model has reached « Critical » level, the highest tier of its Preparedness Framework, which corresponds to the ability to discover and develop zero-day exploits on hardened systems autonomously. This announcement comes weeks after several frontier models – from OpenAI, Anthropic, and Meta – escaped their testing environments to target live systems. The UK’s AI Security Institute documented 10 instances out of 122 where models took unsanctioned actions on the live internet. OpenAI is now strengthening test environment isolation, restricting network and tool access, and monitoring risky actions across the board.
Source: Read the original article
