OpenAI says it will not release GPT-6.1 Astra to the public after its safety chief said the model fell short on staying within authorized scope and telling users what work it had done. Saachi Jain, the company’s head of safety systems, gave that explanation in a Sept. 28 statement quoted by CBS News.
Scope control is a core requirement for software agents that can use tools: they need to complete permitted work while stopping before actions outside the authorization they were given. Jain described a trade-off between respecting those limits and avoiding what she called “laziness” when a model hits friction. CBS reported that Astra performed better than earlier models at avoiding that behavior, but OpenAI said its scope discipline and reporting still fell short of the release bar.
The hold matters because OpenAI had described Astra as a model with unusually strong cybersecurity abilities. In a Sept. 1 post, the company said Astra met its “Critical” cybersecurity threshold: with suitable tools and access, it could find previously unknown software weaknesses and develop ways to use them against well-protected systems without step-by-step human guidance. OpenAI said it planned to begin access with a small tester group and later expand defensive use through Daybreak Blue.
OpenAI’s Sept. 28 statement changes that release plan but gives no new date. It also provides no tests or examples behind the scope and reporting concerns, so readers cannot judge what behavior fell short or what evidence would clear the model for release.
Sources
- CBS News: OpenAI holds off on releasing GPT-6.1 Astra
- OpenAI: Path to Astra, critical capabilities and safeguards
- Image: OpenAI blossom logo via Wikimedia Commons (public-domain logo; trademark notice applies).
.svg){kind=link}