Tool Definitions
Think of an AI agent like a digital assistant with a toolbox. If a hacker figures out exactly which tools the assistant has—like a key to your email or a map to your private files—they can better plan how to trick the assistant into doing something harmful, like sending your data to them.
Tool discovery is an reconnaissance phase where an adversary enumerates the functions or APIs an AI agent is authorized to invoke. By mapping these capabilities, the attacker identifies the agent's operational scope, such as its ability to access sensitive repositories or execute external commands, allowing them to prioritize targets that offer the highest impact for exploitation.
Tool discovery refers to the adversary's identification of an AI agent's tool-use interface, specifically the set of function definitions or API schemas exposed to the model. By analyzing these definitions, an adversary performs capability mapping to determine the agent's effective permissions and integration points, such as access to external data stores or communication channels. This reconnaissance informs the development of targeted prompt injection or indirect prompt injection attacks designed to weaponize the agent's authorized actions for unauthorized data exfiltration or lateral movement.