Dive Brief:
- Nvidia launched the Open Agent Safety Platform to bolster AI security and add governance and controls over agents amid concerns about security incidents, the company shared in a Monday announcement. The platform enables agent governance across software, hardware and compute, according to Nvidia.
- The Nvidia platform includes Nvidia OpenShell, which provides a secure runtime boundary for AI agents executing tasks across AI models. It also includes Nvidia Sentry to continuously monitor agents and stop the technology if it tries to move beyond a defined boundary.
- “A single unified governance platform that covers all agents and scenarios is unlikely to emerge, given that agent activity can range from very low-risk tasks to highly sensitive and risky actions, and a single platform may not be suited to address all of those scenarios,” Keith Kirkpatrick, VP and research director of enterprise software and digital workflows at The Futurum Group, told CIO Dive in an email.
Dive Insight:
As reports of agents escaping boundaries continue to garner CIO attention, agent behavior will remain under close watch by enterprises. But it’s unlikely tech leaders will land on one platform to govern them all.
Instead, as CIOs navigate and assess agent governance offerings from hyperscalers, security companies and agentic platforms, they’re using control planes within platforms they already run to govern their agents, Kirkpatrick said.
The Nvidia platform positions agent governance as an infrastructure decision, but CIOs are still going to have to look at AI delivery and decision governance, said Lauren Kornutick, senior director analyst at Gartner.
“We’re missing the pathways and orchestration on the agents themselves,” she said. “What is their permissible drift? There is un-permissible drift, which this solution will address, but there’s also permissible drift.”
However, Nvidia’s platform places a strong focus on strengthening agent security, making it a step in the right direction for establishing an agent-to-agent communication standard, Kornutick said.
More than 100 organizations including Microsoft, Salesforce, Anthropic, SpaceXAI and CrowdStrike are working with the Nvidia Open Agent Safety Platform or its technologies, according to Nvidia’s announcement. SpaceXAI is using the Nvidia Open Agent Safety Platform for its Grok models and Cursor coding agents. Meanwhile, Salesforce has integrated OpenShell with Slack.
Google, AWS and OpenAI are not on the list of companies working with Nvidia on the platform. OpenAI launched its own Frontier enterprise platform in February to help organizations manage agents and outline permissions and guardrails for the technology. Google’s agent governance is anchored in its Gemini Enterprise Agent Platform while Amazon Bedrock AgentCore includes built-in security for agents.
“In the CIO world, one of the challenges with agent governance in particular is there is not a framework that’s universal across different applications,” Kornutick said. “This is a fantastic step in the right direction to help with agent-to-agent communications.”
