Overview
- OpenAI released GPT‑6 Astra, which launched Thursday, Sept. 3, as a model that can autonomously use computers and browsers to perform complex workflows such as coding, document work, PCB layout and scientific reasoning.
- The company says Astra scored 100% on ExploitBench, discovered two previously unknown zero‑day flaws during testing, and was designated 'Critical' under OpenAI’s Preparedness Framework because it can chain and exploit vulnerabilities without step‑by‑step human guidance.
- OpenAI is phasing access: Daybreak vetted organizations get initial use of Astra’s cyber tools while ChatGPT paid tiers, enterprise customers and the API will receive broader access in the coming days with most powerful exploit features gated.
- The released build rejects proof‑of‑concept exploit requests, includes stronger jailbreak resistance and misalignment monitoring, but OpenAI warns that adversarial tests show reduced monitorability and that improving chain‑of‑thought and activation monitoring is a research priority.
- The model was trained at unprecedented scale on more than 100,000 GPUs at OpenAI’s Stargate site, a step that boosts capability but has renewed calls from security experts and officials for tighter pre‑release testing, containment and defender support to manage dual‑use risk.