OpenAI GPT-6 Astra Reaches 100% ExploitBench but Limits PoC Generation to Patching Only

OpenAI released GPT-6 Astra after it saturated ExploitBench at 100% versus 78.5% for GPT-5.6 Sol. The model also produced higher arbitrary code execution rates on recently disclosed flaws and demonstrated capability against hardened browsers and OS privilege escalation paths when safeguards were removed. The public rollout limits Astra to secure code review and patching workflows only. OpenAI cited its Preparedness Framework critical threshold and added stronger jailbreak resistance plus runtime monitoring. Daybreak for Frontline Defenders commits $1 billion in subsidized access and training for water, power, state governments, and banks, paired with an MS-ISAC pilot. Contract language and procurement records show emphasis on detection engineering rather than offensive tooling. The gap between saturated benchmark performance and released safeguards reveals an internal capability-control split. Prior frontier models showed similar patterns where maximum scores were achieved in red-team conditions but shipped with prompt-level blocks. This creates a window where state or well-resourced actors may seek less-restricted instances before OpenAI expands Daybreak access. Next phase will test whether the $1 billion program produces measurable patch velocity in critical infrastructure or merely increases monitoring overhead. Procurement filings due in Q1 2027 will indicate actual deployment scale versus announced intent.