OpenAI: Astra Achieves Critical Cybersecurity Level

OpenAI has officially announced that its Astra model has achieved a Critical cybersecurity level, indicating its ability to independently identify and leverage zero-day vulnerabilities. Following extensive testing, Astra successfully discovered two previously unknown zero-day flaws within exploit chains and executed complex attack sequences, including escaping sandboxes and escalating privileges to root level. These capabilities demonstrate a substantial advancement in AI's potential for cybersecurity applications.

In response to these developments, OpenAI has enhanced Astra's security measures. The model now refuses to execute prohibited requests in 91.5% of cases, a significant improvement over previous versions like GPT-5.6 Sol. Additionally, internal monitoring and reasoning systems have been integrated to prevent unauthorized actions.

While Astra has not yet been publicly released, OpenAI states that it will be made available soon, with restrictions on the most dangerous functionalities. All testing and benchmarking were conducted internally, and the results remain confidential. This milestone underscores the importance of rigorous safety measures in deploying highly capable AI models in cybersecurity contexts.