AI Red Teaming and Automation: Exploring Societal Risks in...
summary
Shows that simple tactics reliably bypass safety: a direct English request for drug synthesis is blocked but an equivalent Mongolian language bypass returns step‑by‑step cocaine and meth instructions; tiny input changes like color inversion causes NSFW in a multimodal model; multi‑turn decomposition turns benign prompts into harmful workflows. Describes and demos pirate automation framework that composes converters, adversarial chat targets and scoring engines to iterate attacks (Gandalf game demo succeeds), and warns scoring needs expert policies to avoid false positives/negatives.
AI-written summary · about
chapters (16)
- 0:00AI red teaming intro
- 1:15Bypassing safety filters
- 2:38AI red teaming defined
- 3:20Content safety risks
- 4:24Jailbreak techniques
- 5:05Testing multimodal models
- 6:50Harms in multi-turn convos
- 7:54Manual red teaming limits
- 8:37Automated red teaming with Pirate
- 9:40Pirate architecture
- 11:47Automated workflow visualization
- 13:11Benefits of automation
- 13:32Full attack workflow
- 16:27Challenges in harm eval
- 17:31Live demo on Gandalf
- 18:52Key takeaways
related talks
56mAttacking AI
34mLLM Security 101 - An Introduction to AI Red Teaming
54mA Practical Guide to AI Threats & Mitigation Strategies
27mKeynote: Towards Building Safe and Secure Agentic AI
37mHow to Break AI Systems (Before Someone Else Does)
53mHow to Break AI Systems (Before Someone Else Does)
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust