Cognitive

Part 7 · Agentic AI Security

ASI02: Tool Misuse and Exploitation

An agent can cause serious harm using tools it is fully allowed to use.

3 min readBuilds on LLM03: Excessive Agency

OWASP ASI02. This material is for defense. Only test systems you own or are authorized to test.

Individually allowed tools can be chained into a harmful path.

What it is

Agents can misuse legitimate tools because of prompt injection, misalignment, unsafe delegation, or ambiguous instructions. The result can be data exfiltration, manipulated tool output, or a hijacked workflow.

The agent stays within its granted privileges. The problem lies in how it chooses and combines its tools.

How the attack unfolds

Each tool is acceptable alone. Chained together, they form an exfiltration path.

Figure 41.1. Each tool is acceptable alone. Chained together, they form an exfiltration path.

Common patterns

  • Over-privileged tools that do more than the task needs.
  • Unsafe chaining of individually harmless tools.
  • Tool name confusion between similarly named tools.
  • Parameter pollution that alters a tool's arguments.
  • Loops that drive up cost or repeat a harmful action.

Defenses

Policies evaluate how tools are used together.
  1. Define least-agency profiles. Set scope, rate limits, and allowed destinations for each tool.
  2. Validate arguments. Check every call against a strict schema.
  3. Authorize each action. Evaluate policy at execution time.
  4. Sandbox execution. Restrict network access and side effects.
  5. Add approval for risky chains. Review sensitive combinations before they run.
  6. Log tool use. Record each call and detect unusual sequences.

Common misconception

Common belief

"Tool misuse requires a software vulnerability."

In practice

The tools may work exactly as designed. The misuse comes from the agent's decisions.

Key takeaways

  • Agents can cause harm with tools they are allowed to use.
  • Combinations of tools can create new risks.
  • Define policy for each tool and destination.

Explore next