The poop cake: recipes to drive a conversational AI mad
summary
A talk about bypassing safety controls in conversational AI like ChatGPT, showing concrete prompt techniques that can elicit harmful or silly outputs (examples include instructions to commit crimes or a recipe for a feces cake). The presenters review recent developments in large-language chatbot capabilities, demonstrate exploitation of content safeguards, and discuss model biases and the ethical problems these vulnerabilities create.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:02Poop cake origin
- 5:59How language models work
- 12:01Toxic tokens and limits
- 16:20Safeguards and flaws
- 22:24Jailbreak techniques
- 30:10Roleplay and memetic proxy
- 34:30Do Anything Now escalation
- 37:20Bing vulnerabilities
- 41:36Conclusions and risks
- 44:25Closing
related talks
18mDefending the unknown using AI and cybersecurity
13mOf course Dave, I'd love to do that - AI Prompt Injection attacks
55mAI Showdown: The Good, The Glitchy, and The Risky
49mUnderstanding Prompt Injection-Techniques,Challenges and Advanced Escalation
01h 03mKeynote: Red, Blue, and Purple AI
51mChatGPT as a developer's personal servant
06m7 practical tips for safer AI and language models
21mGPT-3 wrote the description for this talk... Scary or exciting?