OpenAI confirms GPT Astra release — the first Critical-level model

🤖 OpenAI confirms GPT Astra release — the first Critical-level model

On September 1, OpenAI published a post titled “Path to Astra”: Astra is the company’s first model to reach the Critical level in cybersecurity under the Preparedness Framework. It independently discovers unknown vulnerabilities and builds exploits without step-by-step human oversight. It scored 100% on ExploitBench, and in a test involving 20 vulnerabilities in the V8 engine, it independently found two zero-days and chained them together, outperforming GPT-5.6 Sol. The release is promised “soon,” with rumors suggesting this week.

🌍 For the first time, a frontier lab has assigned a model the Critical level and disclosed its safety measures: refusals on cyber-jailbreaks — 91.5% versus 59% for Sol, classifiers, CoT-monitoring, and limits for “risky” accounts. A precedent for the entire industry.

👤 AppSec teams and penetration testers will need to rethink their threat models: AI is already autonomously searching for vulnerabilities and building exploit chains.

Source 1: https://openai.com/index/path-to-astra/