OpenAI's new Astra AI model demonstrates significantly improved vulnerability detection capabilities—achieving 100% on the ExploitBench test and identifying two zero-day exploits—while also strengthening refusal safeguards to 91.5% accuracy (vs. GPT-5.6 Sol's 59%). The model will launch soon with restricted cybersecurity features available only to approved organizations. This represents a key advancement in AI safety controls but may limit broad developer access to security testing capabilities.