OpenAI officially launched GPT-6 Astra on Thursday and simultaneously disclosed that the model scored a perfect 100% on ExploitBench — the benchmark that measures a model’s ability to turn known software vulnerabilities into working exploits. Its predecessor, GPT-5.6 Sol, had reached 78.5%. The gap matters: Astra was also able to independently discover and exploit two undisclosed zero-day vulnerabilities found during internal testing between June and August 2026, and could develop privilege-escalation exploits for hardened operating systems when run without safeguards. OpenAI classified the model as reaching the “Critical” cybersecurity capability threshold under its Preparedness Framework before releasing it.
Because the same capabilities that help defenders find weaknesses faster can help attackers exploit them more easily, OpenAI constrained the public release: Astra will help with secure code review and vulnerability patching, but will refuse requests to generate proof-of-concept exploits. The company said it plans to expand access through its OpenAI Daybreak program “in the coming weeks,” with less restrictive safeguards for vetted defensive workflows including vulnerability validation, malware analysis, and detection engineering. To counter misuse, OpenAI added stronger jailbreak resistance, additional monitoring context, and new safeguards to detect and contain misalignment. Astra is rolling out initially to a small set of organizations before reaching all ChatGPT Plus, Pro, Business, and Enterprise users, along with API access via Azure and AWS Bedrock.