OpenAI GPT-6 Astra rollout begins with cyber capabilities restricted

OpenAI has begun the GPT-6 Astra rollout, first providing the model to a limited group of organisations before planned access for paying ChatGPT customers and major cloud platforms. The staged release reflects OpenAI’s designation of Astra as its first model to meet the company’s “Critical” cybersecurity capability threshold, requiring tighter controls over its most advanced cyber functions.

OpenAI said Astra would become available over the coming days to ChatGPT Plus, Pro, Business and Enterprise users, as well as through its API, Microsoft Azure and AWS Bedrock. CNBC reported that organisations in Daybreak, OpenAI’s application-based cybersecurity programme, would receive access first.

The timing remains uneven. Finextra reported on September 4 that some paying users were still waiting for access and cited chief executive Sam Altman describing the process as “messy.” OpenAI’s launch announcement remains the more specific statement on its intended availability timetable.

Who gets GPT-6 Astra first and what can it do?

Initial access is directed at selected organisations, while the broader release is intended for the company’s paid ChatGPT tiers and developer and cloud customers. OpenAI has not said that all users receive the same cybersecurity functions.

According to OpenAI, Astra can carry out computer-based tasks including completing online forms, changing customer information in CRM systems, arranging calendars, researching online and drafting summaries in email or document software. The company also says it can analyse scientific data, build websites, install and test software, and troubleshoot problems visible on screen. These are OpenAI’s descriptions of the model’s intended uses, rather than independently verified performance measures.

OpenAI says Astra also improves performance in browsing, software engineering, science and multistep professional work. Its published benchmark results include a 98% score on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. Those results are reported by OpenAI and use company-selected evaluations.

Why are Astra’s cybersecurity features restricted?

In a September 1 safety update, OpenAI said Astra met the Critical level in its Preparedness Framework. The company defines that threshold as a model’s ability, when given appropriate tools and access, to identify previously unknown vulnerabilities and develop ways to exploit them in many well-protected systems without human guidance at every stage.

OpenAI said the most advanced cyber work would initially be available only to a testing group, with Daybreak Blue later intended to extend defensive use. The restriction is designed to limit both misuse by malicious users and unauthorised actions by the model, according to the company.

The company said it postponed parts of Astra’s development and launch for several weeks to reinforce and test protections against cyber misuse and unauthorised model actions. OpenAI said Astra was not involved in the Hugging Face incident, but that it incorporated lessons from the episode into its safety measures.

OpenAI also said it has added training intended to improve refusals of harmful cyber requests, system protections against misuse and monitoring intended to halt potentially unauthorised activity. It said those safeguards sufficiently reduced the risk of severe harm to release Astra under its framework. That is OpenAI’s assessment, and the company’s phased rollout leaves the most sensitive capabilities under restricted access.

This story draws on original reporting from Finextra Research.