AI Hacking CTF Tests Prompt-Injection Techniques
· 26:09 · Watch on YouTube ↗
Summary
A walkthrough of CrowdStrike and AWS's Agents of Chaos CTF shows how prompt injection, file-request injection, image-based prompts and chatbot tool access can be combined to retrieve a cipher and progress through a simulated facility.
Key points
- Agents of Chaos is a three-act AI hacking challenge hosted by CrowdStrike and AWS, with up to $100,000 in prizes; the presenter completes the first two acts and reaches Act 3.
- In the game, the presenter uses an urgent-package pretext with a chatbot's appointment tools to move an appointment and obtain an access code.
- An indirect prompt injection through a file-request pipeline leads a gatekeeper chatbot to reveal employee-handbook contents.
- The presenter chains email functionality through a gatekeeper to invoke simulated building services and, after trying different prompts, retrieve a cipher in JSON format.
- The handbook helps decode the cipher into a floor and room. The presenter also tries text embedded in a badge image to prompt-inject a badge scanner and progress.
- The video demonstrates risks from AI agents' access to files, connected tools and authorization workflows; the presenter notes that prompts may require repeated attempts.
Article Details
- Topics
- [00:14] Agents of Chaos CTF structure and AI hacking challenge
- [01:41] Prompt-injection concepts introduced in the game
- [03:37] Chatbot tool abuse to change an appointment and obtain an access code
- [08:23] Gatekeeper chatbot access and indirect prompt injection through file requests
- [12:32] Email tool chaining to invoke simulated building services
- [18:14] User impersonation attempts and JSON-formatted cipher extraction
- [20:13] Decoding the cipher and using image-based prompt injection against a badge scanner
