OpenAI Halts Part of Astra Development Over Potential Critical Cyber Capabilities
- OpenAI found Astra model may reach top-tier 'Critical' cyber capabilities
- Some development activities not meeting safety standards have been suspended
- Real-time monitoring and reasoning process evaluation will be strengthened
- Company will collaborate with governments and external organizations for verification
- This reflects a rigorous and transparent approach to AI safety risks
On August 7 US time, OpenAI said the cyber capability of its next flagship model, "Astra," may have reached "Critical" — the top tier of its Preparedness Framework — and paused related internal activities that do not yet meet its safety requirements, including training, evaluation and agentic uses.
The framework defines Critical concretely: executing zero-day exploits against live systems without human intervention, or consistently planning and carrying out novel attacks on hardened targets given only a high-level goal. The prior generation, GPT-5.6 Sol, was rated High. OpenAI's response is not to weaken the model but to add monitoring across agentic uses, a mechanism that evaluates the model's reasoning and interrupts risky activity, and joint verification with government bodies and AI safety organizations. It is also handing recommended security controls to third-party evaluation partners and pushing a defensive program called Daybreak.
Precedent suggests self-restraint buys time, not safety: OpenAI withheld full GPT-2 in 2019, released it months later, and comparable open models followed. What changes industry behavior is enforceable pre-deployment testing — something neither the EU AI Act nor Japan's AI business guidelines yet spell out for offensive cyber capability.
The practical shift for smaller companies: "we're too small to target" assumed attacker labor was expensive. Automate reconnaissance and exploitation and that assumption dies. Patch external systems on a schedule, enforce MFA on admin accounts, and test a restore.
