How to Break AI Systems (Before Someone Else Does)
summary
Shows real exploit chains: hidden prompts in PDFs and emails that exfiltrate password‑reset URLs, multi‑turn “crescendo” jailbreaks that reliably break many models, and poisoning of RAG memory to create persistent data leaks. Demonstrates an open‑source harness called pirate with hundreds of jailbreak templates, scorers and audit memory, reports an RCE in a desktop operator, and recommends layered defenses including spotlighting (delimiting, data‑marking, encoding) plus Guardrails/Nemo prompt filters.
AI-written summary · about
chapters (9)
- 0:00Breaking AI systems
- 2:09AI red teaming and safety
- 11:22Architectural flaws in AI
- 15:18Direct prompt injection
- 22:45Indirect prompt injection
- 30:36Agentic AI vulnerabilities
- 42:33Red teaming tools
- 50:19AI security mitigations
- 1:00:09Q&A
related talks
37mHow to Break AI Systems (Before Someone Else Does)
37mKeynote - Hacking AI How to Survive the AI Uprising
53mHow to Break AI Systems (Before Someone Else Does)
51mPrompt Injection: When Hackers Befriend Your AI
01h 06mHow hacking works
56mAttacking AI
30mTeam Red: Breaking Large Language Models
42mHow to attack AI systems (and how to defend them) !!!!