OpenAI details strengthened controls around preliminary Astra cyber evaluations

OpenAI officially confirmed that preliminary evaluations of the unreleased Astra model meant it could not rule out its “Critical” cybersecurity threshold. It paused internal activities that did not meet strengthened controls and described isolated testing, restricted network and tool access, weight protections, sandboxing and universal monitoring of risky actions. OpenAI separately says GPT‑5.6 Sol remains assessed as High rather than Critical and that Astra was not involved in the reported Hugging Face exploit.

Impact: capability-gated development is becoming a concrete release-engineering practice. The operational actions are primary-source facts; Astra’s capability remains a preliminary, vendor-defined assessment without public model artefacts or independent reproduction.