OpenAI has paused some internal Astra activities after preliminary evaluations said it could not rule out the model reaching its cybersecurity “critical” threshold. The company is treating that as a deployment safety issue, not just a benchmark result, while it continues final evaluation under its Preparedness Framework.
The controls are operational: isolated testing environments, restricted network and tool access, stronger model-weight protection, more monitoring, and sandboxed execution. OpenAI says its monitors review the model’s chain of thought and can interrupt high-risk activity mid-run, which means agentic systems are being handled more like untrusted software than like static models.
That matters wherever an AI agent can use tools on real systems. If the agent can reach code, accounts, or internal services, the exposure is whatever those tools can touch, and the containment stack becomes part of the safety case.