OpenAI announced that Astra, which it plans to make available soon, is the first model to reach critical cybersecurity capability under its Preparedness Framework. The company said the model can identify previously unknown vulnerabilities and develop methods to exploit them. While part of the development and release process has been delayed for security reasons, only certain test users will initially have access to the most advanced capabilities; access will later be expanded with a defense-focused approach through the Daybreak Blue program.
Astra scored 100% in the ExploitBench test. In an internal evaluation covering 20 high-severity vulnerabilities disclosed between June and August 2026, it achieved a higher rate of remote code execution while using fewer tokens than GPT-5.6 Sol, and identified and exploited two zero-day vulnerabilities. In the tests, it escaped the sandbox, executed commands on the host system, and gained root privileges.
OpenAI will strengthen its rejection of harmful requests, account restrictions, and monitoring. While no independent verification is available, the system card and safety evaluations will be shared at launch.
Why it matters
Astra’s capabilities show that AI models are classified not only by their ability to generate content but also by their capacity to find and exploit vulnerabilities in cybersecurity. For this reason, keeping access to the model’s most advanced version gradual and limited has direct implications for both test users and the Daybreak Blue program created for defensive use. However, the fact that the reported performance has not been independently verified leaves open the question of whether the results can be replicated under different conditions. The system card and safety evaluations said to be shared at launch will be essential sources for assessing to what extent safeguards against harmful requests, account restrictions, and monitoring measures are sufficient in the face of this capability.
Background
OpenAI is not a new name in the FikirPilot archive: we have published 7 articles mentioning the name in the last 90 days; the most recent is dated September 1, 2026.