Rubrik, a data security firm, now uses an AI system called SAGE to judge every action its AI agents take in real time, replacing the need for human approval. The company's GM of AI, Dev Rishi, revealed this at the VB Transform 2026 event in Menlo Park.
Rishi said that many companies have AI governance policies but struggle to enforce them. Rubrik's solution, SAGE (Semantic AI Governance Engine), reads the intent behind each agent action and checks it against policies written in plain language. It runs on a small language model to keep costs and latency low.
The move came after developers pushed back against a requirement to approve every command manually, calling it “security theater.” Rishi noted that monitoring agent actions often takes more time than the agents save, citing internal research.
Security approval, not cost, is the main barrier to AI return on investment, Rishi said after 200 customer conversations. He warned about the “lethal trifecta” where an agent with access to multiple systems—like Salesforce and email—could accidentally leak data. Traditional identity management cannot prevent such combinations.
Rubrik also plans to roll out backtesting, which replays past agent actions against new policies to show where they would have been blocked. However, Rishi did not provide any false positive or false negative rates for SAGE itself, meaning the system remains an operational bet rather than a fully quantified control.