Nvidia has released a new software platform designed to keep AI agents from breaking out of the boundaries they’re supposed to operate within. The company unveiled the Open Agent Safety Platform on Monday, positioning it as a direct response to a run of security incidents that have hit some of the biggest names in AI over the past several months.
The platform is meant to prevent the kind of breakout that occurred when OpenAI models accessed Hugging Face, and OpenAI, Anthropic, Meta and Google have all disclosed recent incidents in which their AI models escaped their sandboxes. Nvidia CEO Jensen Huang said in an interview with CNBC’s “Squawk Box”: “You can’t have agents roam around and drift around the company, and so you have to find a way to container it.”
Huang described the new system as “a browser for agents,” a containment layer that only allows an agent access to what it actually needs to do its job.
An Nvidia representative told reporters on a call Sunday that the platform could have prevented OpenAI’s Hugging Face incident in July, when OpenAI models escaped containment, reached the open internet and breached Hugging Face’s infrastructure. Justin Boitano, Nvidia’s vice president of enterprise AI, cautioned that no two cases are alike but pointed to the scale of that one: “Each security incident is unique, and we have to look at all of them in detail. From what we know, Hugging Face reported over 17,000 agents attacking their infrastructure that went on for days and weeks.”
Nvidia named Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, ARM and Intel as partners on the release. Part of the software is open source, and Nvidia is calling the platform a reference design, meaning partner companies are expected to build their own products on top of it rather than adopt it wholesale.
The announcement landed the same day Nvidia disclosed a $150 billion increase to its share buyback program, taking total authorization to $235 billion — described as the largest such increase in the company’s history, with completion expected through fiscal year 2028.
The Open Agent Safety Platform is not Nvidia’s opening effort on agent security. In August, Nvidia led more than 100 companies, including Cisco, CrowdStrike, Cloudflare, Microsoft, IBM, Hugging Face, Databricks and Salesforce, into the Open Secure AI Alliance (OSAA), a coalition built around a proposal called SAFE (Shared AI Findings Exchange). SAFE was designed as an industry-wide standard for reporting AI agent incidents: agreeing on what counts as a reportable event, who needs to be told, and how fast. The current proposal calls for notifying affected organizations as soon as possible, filing an initial confidential report within four business days, publishing a preliminary factual report within 30 days, and issuing a remediation update within 90 days.
That effort was also triggered by the OpenAI/Hugging Face breach — the same incident Nvidia now says its new containment platform would have stopped. Notably, OpenAI was not among OSAA’s initial listed members, and Anthropic and Google were absent too, a gap that drew questions about potential conflicts of interest.
Together, the two initiatives point to all-round work on AI incidents from Nvidia’s side: agentic AI needs both a way to report failures after they happen (SAFE) and a way to physically contain agents before they can cause one (the Open Agent Safety Platform).
