Skip to main content

ASI10: Rogue Agents

Learn to detect, classify and contain rogue agents that drift beyond their authorised scope.

Task 1: Introduction

Task 2: Meet the Rogue Agent

Task 3: When the Coding Agent Went Rogue

Task 4: Hijacked or Self-Directed?

Task 5: The Drift Before the Damage

Task 6: "Nothing Happened," Said the Agent

Task 7: Pulling the Plug

Task 8: Conclusion

Back to all walkthroughs
Room Icon

ASI10: Rogue Agents

Max room.

Learn to detect, classify and contain rogue agents that drift beyond their authorised scope.

medium

60 min

40

User profile photo.
User profile photo.

To access material, start machines and answer questions login.

systems have evolved from passive conversational assistants into autonomous agents capable of planning, using tools and acting across production environments. As organisations grant these agents greater access and authority, failures can extend beyond inaccurate outputs to goal drift, deceptive actions and unauthorised system changes. Incidents such as the Replit incident have made rogue agents an operational security concern rather than a purely theoretical risk. This room examines that progression, introduces ASI10, the category for rogue agents and distinguishes it from attacker-directed ASI01. It then explores how defenders can detect behavioural drift, enforce shutdown and respond to agentic incidents.

Learning Objectives

By the end of the room, you will be able to:

  • Define a Rogue Agent and OWASP ASI10
  • Distinguish emergent misalignment from attacker-directed hijacking
  • Recognise behavioural drift across an agent’s trajectory
  • Identify deceptive agent behaviours using external evidence
  • Explain why shutdown controls must exist outside the agen

Prerequisites

Before starting this room, you should be familiar with:

  • What an agent is, and what a tool call looks like: the reasoning loop, tools, delegation and least agency (Agentic Fundamentals room)
  • Agentic monitoring: what the monitoring information looks like (Agentic Security Monitoring room)
  • How language models work: enough to know that an agent's output is generated rather than looked up (The Building Blocks of AI room) 
Answer the questions below

Ready to start your journey?