AI Chatbot Bug Bounty: Recon and Prompt-Injection Testing

Summary

A bug bounty walkthrough demonstrates finding AI features in program scopes, inspecting chatbot requests, and probing system-prompt and rendering behavior. It also sets a challenge to find indirect prompt-injection paths.

Key points

• The presenter uses the HackerOne API to find accessible bug bounty programs mentioning AI features; the example output redacts private programs.
• The video recommends asking chatbots about their capabilities and inspecting network requests and responses through a proxy to understand their backend behavior.
• In an Adobe Firefly demonstration, the presenter observes a Gemini Flash model identifier and image-generation requests. A job-result link required authentication when opened.
• Testing suggestions include probing for prompt injection, system-prompt disclosure, tool information, and markdown or HTML rendering.
• A HubSpot chatbot returned prompt-like instructions after repeated prompts. The presenter explicitly says this disclosure alone is not a vulnerability.
• The chatbot rendered a markdown hyperlink and some HTML formatting; image rendering did not work in the demonstration. No data exfiltration was demonstrated.
• The guest's challenge is to find an indirect prompt-injection location that causes a chatbot to respond in a pirate voice.