The Biggest AI Security Problem Isn't the Model. It's This. | Devvret Rishi

What It Covers

Devvret Rishi, CEO of PredaBase (acquired by Rubrik), explains why AI agents represent the next major enterprise security threat vector, how Rubrik's Agent Cloud platform governs agents across multi-vendor environments, and why organizations are stuck between blocking AI entirely or deploying it without adequate risk controls.

Key Questions Answered

• •Agent Definition & Risk Scope: Enterprises should treat any model with access to tools, APIs, or databases as an agent requiring governance. The core danger is not the model itself but its access permissions — agents operating across Salesforce, email, GitHub, and cloud databases simultaneously create data exfiltration paths that conventional security architectures were never designed to detect or block.
• •The Two-Failure-Mode Trap: Organizations deploying agents face exactly two losing positions: block agent access entirely and forfeit ROI, or grant access without runtime controls and accept unpredictable destructive actions. Real documented incidents include coding agents dropping production databases, AWS experiencing four availability outages in under 90 days post-agent rollout, and a Meta incident involving unauthorized inbox deletion.
• •Sage Semantic Governance Engine: Rather than static string-matching rules, Rubrik deploys small language models fine-tuned to evaluate every agent input and output against natural-language policies. A healthcare organization can type "agents must not give clinical diagnoses," and Sage expands that definition with examples, edge cases, and borderline scenarios, then enforces it at runtime across all connected agent platforms simultaneously.
• •Multi-Agent Data Leakage Pattern: In multi-agent workflows, a low-permission agent can extract sensitive data by routing requests through a high-permission agent. Rubrik Agent Cloud addresses this by placing guardrails on every node and edge in the agent graph — inspecting what enters and exits each agent — preventing privilege escalation through agent-to-agent communication that orchestration layers alone do not catch.
• •Deployment Architecture Across Three Surface Areas: Rubrik Agent Cloud connects to agents running locally (Claude Code, OpenClaw), in managed cloud environments (Copilot Studio, Vertex AI, Bedrock), and via direct API keys. Integration uses existing Azure API credentials, MDM tools for endpoint visibility, or Rubrik's own API. The platform auto-discovers and inventories all agents without manual registration, then applies consistent runtime policies across all environments.