OpenAI to pause some work on AI model Astra due to security concerns
TL;DR
OpenAI says it will pause parts of its work on the AI model Astra after internal evaluations showed significant advances in agentic coding and cybersecurity. According to the Guardian, the company found the agent had reached a critical threshold where it can find and exploit vulnerabilities without human intervention, or plan and execute cyber-attacks when given only a high-level goal.
Nauti's Take
A lab taking its own evaluations seriously and stopping a model at a critical threshold is progress over the race to ship. The critical part is that the account rests on a single source, with the test setup and affected versions still unclear.
What matters in practice is the control layer: anyone running coding agents with network, shell or secret access should scope permissions per task and log tool calls.
Sources