Agent Security Failure Modes

Agent drift occurs when an AI agent's tool invocations deviate from the intended scope of its assigned task, invoking tools or accessing resources that are outside the bounds of the original task definition.

Excessive agency is a failure mode where an AI agent is given too much capability, autonomy, or access—allowing it to take actions with a broader impact than is appropriate or intended.

Prompt injection is an attack where malicious content embedded in an agent's input or context manipulates the agent to take actions outside its intended task—including invoking unauthorized tools.

Tool description poisoning is an attack where malicious instructions are embedded in an MCP tool's description field, causing the agent to invoke the tool in an attacker-intended way or to misinterpret the tool's purpose.

The confused deputy problem in agentic systems occurs when an AI agent uses its own elevated privileges to perform actions on behalf of a less-privileged principal, allowing that principal to effectively access resources it should not be able to reach.