OpenAI says its upcoming Astra model may be its first to approach the top rung of its own cyber-risk scale, and it has changed how it secures the model while testing continues.

In a post published Aug. 7, 2026, OpenAI reported that recent internal evaluations showed Astra making measurable gains in agentic coding and cyber tasks. The company concluded it “cannot rule out critical cyber capabilities,” a step above the “High” tier assigned to its current flagship, GPT-5.6-Sol.

Under OpenAI’s Preparedness Framework, the “Critical” tier describes a model that can autonomously discover and weaponize unknown vulnerabilities across many hardened systems, or plan and execute a full attack chain against a hardened target from a single high-level goal. OpenAI stressed the finding is preliminary and it has not confirmed Astra crosses that line. OpenAI did not say which criterion the results triggered.

In response, OpenAI updated its safeguards to include isolated test environments, restricted network and tool access, stronger model-weight encryption, added monitoring, sandboxed execution, continuous review of the model’s reasoning during agentic runs, and a pause on internal work that does not meet those controls. It also plans to involve government agencies and outside safety organizations in testing.

For vendor-risk reviews, this is a self-reported assessment, with no independent verification, and no benchmark scores disclosed. Still, the named controls give security teams a practical checklist to require from vendors before approving frontier models for coding or agentic workflows.

Source: OpenAI