# AI Runtime Protection Secures AI Agents by Monitoring Prompts and Tool Calls

This talk explains how organizations can secure AI agents at runtime, with a focus on where to place controls and what those controls must evaluate. Tal Bronfer outlines Noma Security’s broader platform—discovery, runtime protection, and red teaming—but centers on runtime defense for prompts, responses, tool calls, and tool outputs so agents do not “go rogue.” He defines agentic control points as the places between an agent and the outside world, especially the inference gateway that sees prompts and proposed actions, the MCP or action gateway that sees tool execution, and endpoint or harness hooks used for local agents. A key point is that no single control point is enough because enterprises run different agent families, including local, homegrown, and SaaS agents. Using a support-agent example, he shows how indirect prompt injection can cause sensitive CRM notes to be exposed, illustrating the “lethal trifecta” of private data access, untrusted content, and outbound actions, and argues that strong runtime protection must combine entitlement checks, contamination detection, and session-level intent analysis.