Of course Dave, I'd love to do that - AI Prompt Injection attacks
summary
Shows concrete attacks against ChatGPT and OpenAI APIs that make models follow hidden commands: direct prompt injection, indirect prompt injection, multimodal injection. Demonstrations include a moderator bypass where appended instructions make the model blame the wrong user, a log-analyzer forced to reply only "false", a fetched webpage that turns responses into "potato", an image prompt that hides people, and simple scripts that turn models into Twitterbots or CV-generators that reliably boost applicants. Some models resist; others obey.
AI-written summary · about
chapters (10)
- 0:00Prompt injection attacks
- 1:30System and user prompts
- 3:45Direct vs indirect injection
- 4:27Direct injection examples
- 6:16AI logs and injection risks
- 7:40Indirect injection in fetched data
- 9:06Multimodal injection with images
- 10:36Corporate chatbot failures
- 11:40AI filtering CVs and hidden instructions
- 12:44Q&A
related talks
24mSecuring Language Models Against Prompt Injection
51mPrompt Injection: When Hackers Befriend Your AI
43mWords as weapons: The dark arts of Prompt Engineering
49mUnderstanding Prompt Injection-Techniques,Challenges and Advanced Escalation
01h 03mKeynote: Red, Blue, and Purple AI
30mBattling the Machine: Cutting-Edge Attacks and Defenses in Generative AI
25mAttack, Defense & Danger in the Age of AI
18mDefending the unknown using AI and cybersecurity