OpenAI: Astra can autonomously discover zero-day vulnerabilities and develop working exploits. OpenAI stated that the upcoming Astra model has achieved a major milestone in cybersecurity, becoming the first model the company has classified as “Critical” capability under its Preparedness Framework. To reach this level, a model must be able to autonomously identify previously unknown software vulnerabilities, develop working exploits on hardened real-world systems without step-by-step human guidance, or carry out an attack starting from a high-level objective alone. During testing, Astra: Achieved 100% on the known-vulnerability exploit development benchmark Discovered two previously unknown vulnerabilities while constructing an exploit chain Escaped a hardened browser sandbox and executed commands on the server Combined multiple operating system vulnerabilities to gain root access OpenAI also noted that Astra outperformed GPT-5.6 Sol in a test designed to detect whether models use prohibited “shortcuts” when faced with extremely difficult or nearly impossible hacking tasks. Astra did not use these prohibited methods but still successfully completed several tasks legitimately. Due to the associated risks, OpenAI has delayed certain aspects of Astra’s development to incorporate additional safeguards and initially plans to provide its most advanced cybersecurity capabilities only to a selected group of testers. This development is particularly significant for the crypto industry, where a single software vulnerability can quickly translate into direct financial loss. More powerful AI models could reduce the time required to discover vulnerabilities and develop exploits from days or weeks to machine speed.