Interrogators: Attack Surface Mapping in an Agentic World

This research introduces the concept of AI agent interrogators and the open-source project Agent Interrogator, an opaque box interrogation framework designed to map the attack surface of agentic systems.

As the adoption of AI agents rapidly expands, there is a growing need to develop the ability to map the attack surface behind their natural language interface, which traditional security tooling cannot accomplish. Through Agent Interrogator, a two-stage AI-assisted interrogation process is employed. Identifying the agent's high-level capabilities in the initial interrogation stage and then enumerating the supporting invokable tools for each capability.

This research validates the approach against test targets utilizing LangChain and Model Context Protocol (MCP) to deliver agentic capabilities. The product of the interrogation is a structured profile mapping the agent's attack surface, enabling security practitioners to identify vulnerabilities such as excessive agency and conduct targeted fuzzing. This work provides a critical foundation for securing the next generation of AI systems and the development of automated attack surface mapping in complex, multi-agent ecosystems.