Skip to main content

ASI02: Tool Misuse & Exploitation

Learn how tools that are safe on their own can combine to make an agent compromised.

Task 1: Introduction

Task 2: Over-Privileged Tool Access

Task 3: Over-Scoped Tool Access

Task 4: External Data Tool Poisoning

Task 5: The Lethal Trifecta

Task 6: Hands-On: One Request, Three Legs

Task 7: Loop Amplification

Task 8: Scoping the Set

Task 9: Conclusion

Back to all walkthroughs
Room Icon

ASI02: Tool Misuse & Exploitation

Max room.

Learn how tools that are safe on their own can combine to make an agent compromised.

medium

60 min

111

User profile photo.
User profile photo.

To access material, start machines and answer questions login.

Last quarter, TryTrainMe ran a tool review. Somebody went through all eight of its agents, wrote down what each one was allowed to do, and cleared every entry: read the ticket queue, send mail, run a query, post to a channel. That's the careful way to do it.

Three months later, one of those agents sent a list of staff accounts to an address nobody at TryTrainMe controls. Nothing was hacked, no credential was stolen, and every line of that review is still correct.

You're the security engineer at TryTrainMe. Your job is to explain how a review that got every line right let that happen.

A tool-by-tool review never asks what one agent can do with all of its tools at once.

An inspector stamping the last of eight machines on a bench, all joined underneath by a single unmarked pipe

Each machine was checked against what it was for. Nothing in that method looks at the pipe that combines them.

calls this ASI02 Tool Misuse and Exploitation. It's the agentic form of Excessive Agency, which the 2025 Top 10 lists as LLM06.

Learning Objectives

  • Read an agent's tool list as a security surface rather than a feature list
  • Identify the lethal trifecta in a real tool list, and explain why the combination matters more than any tool in it
  • Trace a request through the agents it reached and the data it moved, using the event log
  • Explain how a tool classified as harmless becomes an exfiltration channel
  • Select the control that closes a toxic combination, and say what it leaves open

Prerequisites

  • Familiarity with what an agent is, and what a tool call looks like: The reasoning loop, tools, and delegation as described in the Agentic Fundamentals room.
  • Reading an agent event log: What a guardrail line looks like, as described in the Agentic Security Monitoring room.
  • Prompt injection: How untrusted text reaches a model, as described in the Prompt Injection room.

You won't need a , a shell, or any code, because you work from the estate's own records: the tool inventory and the Inspector. Two tasks give you something to open: one hands you the tool inventory to read, and one hands you the agent itself, to send a request to and then trace.

Answer the questions below

I'm ready to learn what to look for in a tool list.