- Anavem reports OpenAI published a designation and system card saying Astra discovered two zero-day vulnerabilities in evaluation.
- The 'Critical' threshold is defined by the model's ability to find and exploit novel security flaws across hardened systems without step-by-step human guidance.
- OpenAI said it is in the process of disclosing those vulnerabilities to maintainers.
- Astra is off by default in Business and Enterprise workspaces and requires deliberate admin enablement.
- OpenAI's monitoring can slow, pause or stop tasks (including legitimate defensive work) when safety checks trigger.
- API usage reportedly exposes new failure modes where long-running agent tasks may be stopped mid-run.
Useful?