OpenAI's Astra model hits highest cyber risk tier, deemed safest yet

openai's astra model is classified as the company's most dangerous model yet, hitting the highest risk tier for cybersecurity in openai's preparedness framework. it demonstrated capabilities to find and exploit previously unknown security holes in well-protected systems without human guidance. astra scored full marks on exploitbench, a benchmark measuring how well a model builds exploits from known vulnerabilities. it also found two previously unknown zero-day flaws and chained them into a working exploit. openai claims astra is also its safest model, with 91.5% refusal of disallowed cyber requests in internal evaluations. the company has implemented new safety measures, including classifiers to watch for potential misuse in production.