Security vendor bets on semantic AI to tame agent risk
A governance gap is stalling enterprise AI deployments - and data security vendor Rubrik says it has a fix built on a custom language model.
The company has launched SAGE (Semantic AI Governance Engine) at RSAC 2026 in San Francisco. Rubrik describes it as the data security industry’s first AI governance engine for controlling autonomous agents in real time.
SAGE forms the core of Rubrik Agent Cloud, the company’s platform for deploying and auditing AI agents at enterprise scale. It is driven by a proprietary Small Language Model (SLM) that Rubrik developed in-house.
The underlying problem SAGE targets is well recognised. Existing enterprise governance frameworks rely on deterministic, keyword-based rules. Those rules cannot interpret natural-language instructions or respond dynamically when an AI agent takes an unexpected action.
Rather than matching keywords, SAGE interprets the intent behind a written policy. An instruction such as “Do not give financial advice” is converted into machine-enforceable logic that the engine can apply to ambiguous or novel agent behaviours.
Announced capabilities include adaptive policy improvement, which Rubrik says flags ambiguous guardrails before a violation occurs, and a remediation feature called Agent Rewind. Rubrik claims Agent Rewind can instantly undo destructive agent actions and restore data integrity.
Rubrik conducted an internal benchmark comparing its SLM against OpenAI’s GPT-5.2 for policy violation detection. The company reported its model processed messages five times faster and detected violations more accurately.
"SAGE marks a pivotal moment in AI security as we shift the focus from if agents can be deployed to how they can be governed at scale," said Devvret Rishi, General Manager AI, Rubrik.
"With SAGE, we can move beyond simple monitoring to a future where AI helps us govern AI agents. Now, we give CISOs the guardrails they need to let their AI agents run at full speed without compromising the security and integrity of the enterprise."