OpenAI Skips Nvidia's Open Agent Safety Platform, Works With Chipmaker Behind the Scenes

Nvidia has rolled out an industry coalition aimed at stopping AI agents from going off the rails, and one name is conspicuously missing from the public backer list: OpenAI. The ChatGPT maker is not among the companies publicly endorsing the Open Agent Safety Platform, but it is working with Nvidia on the effort behind the scenes, according to reporting by TechCrunch. The split between quiet cooperation and public silence is the story here, because agent safety is shaping up as the next front where labs, chipmakers and regulators stake out positions.

The platform, as Nvidia frames it, is meant to give companies shared tools for keeping autonomous agents inside bounds. Agents are the software layer that takes a goal and executes multi-step tasks on its own, booking, buying, coding and calling other systems without a human approving each move. That autonomy is exactly what makes them useful and what makes them hard to contain. An agent with access to a company's internal tools can cause damage quickly if it misreads a goal or gets manipulated by a malicious prompt.

Nvidia's position in the debate is unusual because it sells to nearly everyone. Its GPUs underpin the training and serving of models from OpenAI, Google, Anthropic, Meta and a long list of smaller labs, so a safety framework backed by the chipmaker carries weight across the field. Getting those rivals to agree on common guardrails is a diplomatic task as much as a technical one, and the public roster of supporters signals which companies are willing to be counted.

OpenAI's absence from that roster does not mean it is sitting out the technical work. The company has its own safety and preparedness teams and has published frameworks for evaluating risky model behavior, including agentic systems. Working with Nvidia privately lets OpenAI shape the technical direction without tying its brand to a framework it may not fully control. That is a familiar posture for frontier labs, which often cooperate on standards while competing hard on products.

For enterprises, the practical question is whether these efforts produce anything an IT team can actually deploy. Agent deployments are already moving into customer support, software engineering and back-office operations, and buyers want to know how to log what an agent did, cap what it can spend and shut it down mid-task. If Nvidia's platform delivers shared specifications for those controls, it could become the default plumbing vendors build against, with or without a logo from OpenAI on the page.

What to watch next is whether OpenAI eventually signs on publicly and whether other holdouts follow. Nvidia has not said how the platform will be governed, who sets the rules or how compliance will be checked, and those details will determine whether this is a real standard or a marketing umbrella. The answer matters beyond the two companies: agent safety rules written now will shape what businesses are willing to let autonomous software do.