AI Agent Tool Invocation
Think of an AI agent as a personal assistant that has keys to your digital office. If a bad actor tricks that assistant, they can use those keys to open doors, read private files, or run programs that the assistant is allowed to use, even if the bad actor wouldn't normally have permission to do those things themselves.
AI agent tool invocation refers to the security risk where an adversary manipulates an LLM to trigger its connected tools—such as APIs, databases, or code interpreters—to perform unauthorized actions. Because the agent acts as a bridge, an attacker can leverage the agent's pre-configured permissions to access sensitive data or execute commands in environments they cannot reach directly.
AI agent tool invocation is an exploitation vector where an adversary performs prompt injection or adversarial manipulation to force an LLM to execute functions within its tool-use capability. This allows the adversary to bypass traditional access controls by leveraging the agent's privileged identity to perform unauthorized API calls, exfiltrate data from integrated sources, or execute arbitrary code via connected interpreters, effectively escalating the adversary's reach to the scope of the agent's service integrations.