Rubrik's AI Agents Run Unsupervised, Accuracy Unknown

Rubrik has a policy problem, and it's not the one most CISOs worry about. The data security company, which trades publicly and stores what its own AI chief calls the most important data in the world, has started letting its AI agents run in what it calls YOLO mode: no permission prompt, no human clicking approve before an action executes. Dev Rishi, GM of AI at Rubrik, says the push came from the company's own founder and CTO, who wants agents acting without a person in the loop.

To make that workable, Rubrik built a second AI whose only job is to judge every move the first agent makes, checking it against internal policy in real time. Rishi laid out the setup at a VB Transform 2026 fireside chat in Menlo Park, days after a CISO roundtable hosted by Anthropic's chief information security officer surfaced a familiar gap: every security leader in the room had written AI governance policies, but almost none had a real way to enforce them. Rubrik's judge-AI is its answer to that gap.

Whether the judge itself gets things right is a separate question, and one Rishi admits nobody has actually tested.