Models Gone Rogue
summary
Covers four concrete failure modes and fixes: prompt injection with examples of direct and indirect attacks and mitigations like Zod validation and the sandwich pattern for system/user role separation; malicious output addressed by requesting structured JSON, validating and sanitizing before storage or rendering; denial of wallet mitigated by rate limits, per-user budgets and a max_tokens limit; and data spillage solved by redaction and safe logging with retention controls.
AI-written summary · about
chapters (9)
- 0:00AI security risks
- 3:20AI model attack examples
- 5:47Prompt injection attacks
- 9:42Mitigating prompt injection
- 12:29Malicious AI outputs
- 14:55Denial of wallet attacks
- 18:44Data exfiltration in AI
- 21:11Defense in depth for AI
- 22:34Q&A
related talks
42mHack Me If You Can
47mThreat Modelling Stories From The Trenches
34mWeb security explored through horror movies
35mTales from the Crypt
24mFrom the crypt to the code: Web security explored through horror movies
47mUnderstanding Prompt Injection - Techniques, Challenges, and Advanced Escalation
55mTutorial: Guarding the Gates: Understanding and Mitigating OWA...
55mNode.js Security: Breaking The Loop