Buried in OpenAI’s Astra preparedness write-up are the evaluation details security teams will screenshot. On public ExploitBench, Astra hit a perfect 100% developing exploits from known vulnerabilities. Worried about contamination, OpenAI built an internal “ExploitBench – Internal Port” with 20 high-severity V8 issues disclosed more recently (June–August 2026 window); there Astra achieved much higher arbitrary code-execution rates than GPT-5.6 Sol while using far fewer output tokens.
The company also describes chain-capable exploit behavior and new refusal/monitoring layers meant to stop real-world abuse prompts. None of this is a consumer feature list — it is the evidence pack OpenAI is using to justify both the Critical label and the decision to bottle the sharpest tools behind partner programs. Read it as a capability disclosure, not a how-to.
Source: https://openai.com/index/path-to-astra/