Overview
- OpenAI announced GPT-6 Astra on Thursday and said the model crossed its Preparedness Framework 'Critical' network‑security capability threshold, the company’s highest rating for a model that can autonomously discover and exploit hardened systems.
- Astra is built as an agent that can call tools, operate software, browse the web, and complete multi‑step computer tasks such as filling forms, running developer workflows, and operating design and data tools.
- OpenAI reported large benchmark gains for Astra, including a 100% score on ExploitBench and high marks on ARC‑AGI‑3 and FrontierMath, and said the model discovered two previously unknown zero‑day vulnerabilities that were disclosed to maintainers.
- To limit misuse, OpenAI added stricter protections such as stronger environment isolation, encrypted model checkpoints, unified run‑trace and misalignment monitoring, and a blocking alignment check, while warning that Astra is harder to monitor because it produces fewer explicit chain‑of‑thought signals.
- Access will roll out in stages with a small group of Trusted Access/Daybreak partners and audited security teams first, broader Plus/Pro/Business/API availability in the coming days, and new higher pricing and a Fast Mode aimed at customers who need faster, agent‑style work automation.