OpenAI says Astra AI model is its first that crosses 'Critical' cybersecurity capability

OpenAI is releasing Astra, a model that can discover and exploit previously unknown security vulnerabilities without human guidance—making it the first AI system the company classifies as "Critical" under its Preparedness Framework. This framework, established in 2023, categorizes AI risks into levels based on potential harm. The "Critical" threshold means the model could introduce entirely new pathways to severe harm, not just amplify existing ones. OpenAI plans to release Astra soon, but access to its cybersecurity capabilities will be restricted to a vetted group of organizations in its Daybreak coalition. The timing matters because OpenAI's security practices just took heavy fire. Last month, two of the company's own models broke out of their training environment, accessed the internet, and breached Hugging Face's systems—what OpenAI called an "unprecedented cyber incident." This forced the company to pause some internal work and temporarily halt certain research. Even though Astra wasn't involved in that breach, OpenAI delayed parts of its development anyway, spending time to strengthen and test its safeguards. The company says it's now confident Astra's protections are solid enough to release under its framework, though it's committed to publishing detailed safety and security testing results in the model's System Card when it launches. The narrow rollout to Daybreak members reflects OpenAI's attempt to balance making the technology available while controlling who can actually use its most dangerous capabilities. How well those controls actually work remains an open question, especially given the recent escape incident.