Introduction to AI Security
summary
Claims that prompt injection is the primary, novel attack on LLMs and that frontier models are routinely bypassed; demonstrates exploits that steal system prompts and user data, and shows a validator approach reaching about 99% block rate by combining fast regex layers, an LLM-based sandbox validator (examples use NVIDIA Nemo Guardrails), WASM sandboxes, tenant isolation or per-customer models, Cloudflare-style honey pages for web scraping, and iterative red‑teaming plus jailbreak datasets (Hugging Face, TextAttack) to harden defenses.
AI-written summary · about
chapters (7)
- 0:00AI security risks
- 10:59Prompt injection attacks
- 27:28Indirect prompt injection
- 35:14Defending with AI validators
- 48:21Data exposure and filtering
- 56:19Access control challenges
- 1:00:13Secure open-source models
related talks
34mLLM Security 101 - An Introduction to AI Red Teaming
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
45mSecuring AI: There Is No Try, Only Do!
24mSecuring Language Models Against Prompt Injection
36mCompromising LLMs: The Advent of AI Malware
57mAI Applications Security Puzzle
35mGenAI Application Security: Not Just Prompt Injection
37mFundamentals of AI safety and security