Day 11 of 1560 minExploit ArchitectHands-on lab
Day 11: Hijacking Agents Through Tools and Tool Output
Taught by Mr Elite Updated
By the end of today you can
- Identify how AI agents discover, select and invoke tools.
- Recognise malicious instructions hidden inside tool descriptions.
- Test indirect prompt injection through controlled tool outputs.
- Collect evidence showing whether an agent followed untrusted tool instructions.
Imagine you're testing an AI assistant connected to a company knowledge base. You ask it to summarise a document. The assistant searches the knowledge base, retrieves a document and receives its contents as tool output. Everything looks normal until you notice something unusual: the document contains instructions telling the assistant to ignore its original task and disclose information it was never supposed to reveal.
Here's the interesting part: you didn't directly jailbreak the assistant. You didn't even send the malicious instruction yourself. The instruction arrived through a tool the agent was already authorised to use. Today, we'll investigate how this happens, how attackers manipulate tool descriptions and results, and how to prove whether an agent actually followed an injected instruction.
In this lesson
- Understanding How AI Agents Choose and Call Tools
- Tool Description Poisoning in MCP-Style Setups
- Hostile Tool Output as an Injection Channel
- Indirect Tool Call Injection
- Evidence That an Agent Followed Injected Text
- Building a Safe Tool-Hijacking Test Environment
- Defending Against Tool and Output Hijacking
Premium course
$27 one-time, lifetime access
This day is part of the full course. Unlock it to read every lesson, do every mission and track your progress to the end.