OpenAI: Astra is first model at Critical cybersecurity threshold

· opgehaald 09:13

OpenAI says upcoming model Astra is its first to meet the Critical cybersecurity threshold under the Preparedness Framework. A public release is planned "soon," but advanced cyber capabilities will be limited to Daybreak partners.

On 1 Sep 2026 OpenAI published Path to Astra: Astra is the first OpenAI model to reach Critical cyber capability — finding previously unknown flaws and exploiting them across well-protected systems without step-by-step human guidance. OpenAI still plans to make Astra available soon, but its most advanced cyber tools stay limited to Daybreak partners (Daybreak Blue / more restricted Daybreak Red). The post cites a perfect ExploitBench score and two zero-days in a modified test; safeguards include stronger refusal training, higher-risk account limits, and extra chain-of-thought monitoring. Full System Card details come at launch.