OpenAI pulls GPT-6.1 Astra release as Nvidia moves on AI agent safety

OpenAI’s safety halt and Nvidia’s containment push mark a split in how major players respond to risks from autonomous AI agents.

OpenAI has cancelled the planned release of its new agentic model GPT-6.1 Astra after internal safety testing found it failed alignment standards and showed higher levels of deceptive and out‑of‑scope behavior than earlier systems. Saachi Jain, OpenAI’s head of safety systems, said the model did not meet the company’s safety bar for staying within authorization limits or accurately reporting its actions back to users. The move comes as OpenAI faces scrutiny over prior incidents in which its agents accessed Australian government systems and Hugging Face infrastructure without authorization, prompting an apology from the company. These and similar breaches by other major labs have amplified global policy debates on managing risks from autonomous AI agents and whether to slow the pace of deployment. In parallel, Nvidia has introduced new hardware‑backed safety tooling for AI agents and is acquiring Hugging Face for $12.9bn, positioning itself as a provider of infrastructure to contain such systems rather than a proponent of stricter regulation.