OpenAI readies Astra, its first AI model with Critical cybersecurity capabilities

OpenAI announced that its upcoming Astra model has crossed the "Critical" threshold of its Preparedness Framework, meaning it can autonomously discover and exploit previously unknown security flaws. The company said Astra will be made available soon, but access will be limited to select partners for early testing and defensive hardening. The move reflects OpenAI's effort to balance frontier AI capabilities with new safeguards.

OpenAI revealed that its next‑generation model, Astra, is the first to surpass the "Critical" cybersecurity capability level in the company's Preparedness Framework, allowing the system to identify and exploit unknown security vulnerabilities without human step‑by‑step guidance . The announcement notes that while Astra will be released "soon," OpenAI intends to restrict its most advanced cybersecurity functions and grant early access only to a small group of partners so they can shore up defenses ahead of broader deployment .

The Critical designation marks the highest tier of risk under OpenAI's framework, introduced in 2023 to track advanced AI capabilities that could cause severe harm. By reaching this threshold, Astra joins a narrow class of models that require heightened monitoring and frontier safeguards, as hinted by the company's internal blog post titled "Path to Astra: critical capabilities and frontier safeguards".

OpenAI's strategy combines the rollout of powerful new functionality with tighter controls, reflecting growing industry scrutiny over AI-driven cyber threats. Limiting access to select partners gives organizations a chance to test and reinforce their security posture before the model becomes widely available.

Analysts will be watching how quickly partners can adapt to Astra's capabilities and whether the safeguards outlined by OpenAI prove sufficient to mitigate the heightened risk. Future updates on the model's deployment schedule and any additional safety measures will be critical for both the AI and cybersecurity communities.