Can new tools help organizations lock down “rogue” AI agents?

Sponsored by

AI agents are starting to do more than access systems. They can invoke tools, execute workflows, and take action across business applications. As that autonomy grows, IT needs visibility into what each agent is allowed to do, what it actually did, who owns it, and how to revoke access when needed. Learn how JumpCloud helps govern AI agents alongside human and device identities.

More software and hardware vendors are offering up new tools to help prevent “rogue” AI agents from breaking free and wreaking havoc online. But will these solutions be enough for IT professionals to effectively protect their organizations?

At the end of September, Nvidia announced OpenShell, an open-source runtime explicitly designed to keep agents from unauthorized system and data access. In addition to that, the company unveiled Sentry, an authentication tool that carefully monitors agents’ requests and responses.

Arjun Bhatnagar, CEO and co-founder of digital privacy startup Cloaked, told IT Brew that Nvidia’s approach to securing agents is “definitely work that’s needed to be done,” while adding that much will come down to execution.

Integrating these types of solutions into organizations is a big question, Bhatnagar added: “It’s about to become a lot more work, and I don’t think it’s going to be with, as we’re hoping, a one-click solution.”

How it works. Ali Golshan, senior director of product at Nvidia, told IT Brew that, as his team observed the types of actions causing agents to go rogue, they realized those agents are often working in environments not designed for them.

“It’s like when we used to try to run monolithic on-prems on top of microservices and Kubernetes; it’s just not designed for it,” Golshan said. “We started thinking about, ‘What does an agent-native software stack look like?’”

While IT pros have been examining whether to put agentic controls, governance, and enforcement at the application model harness layer, Golshan pointed out that agents have gotten more powerful and capable of working together. That means a more comprehensive approach to agentic governance and guardrails, including the runtime layer.

To that end, OpenShell is meant to ultimately allow an organization or user to “draw a deterministic line” so that the agent “cannot reason or have a choice about how they operate or behave,” Golshan said.

Bhatnagar suggested that OpenShell is a step in the right direction, but not a holistic solution to the security issue surrounding AI agents.

“You’ve now given people the opportunity to start working on the problem…versus fixing it with this one solution,” Bhatnagar said.

Solving the rest of the problem. According to Bhatnagar, managing agentic security is best approached through establishing infrastructure and then having a team play with agents in a controlled environment, like a sandbox, instead of allowing organizational-wide agentic access on collaborative messaging platforms like Teams or Slack.

Once the IT department figures out how employees interact with agents (including the data they share), they can set the proper data policies and controls.

“As an IT professional and security leader, you have the rare opportunity to have executive presence on helping the business leaders try to embrace and unlock business productivity use cases,” Bhatnagar said. “See where the agents apply with your corporate data.”