Securing Language Models Against Prompt Injection
summary
Lists three concrete attack classes: prompt injection (examples include “ignore all previous instructions” and role‑play do anything now or “finish every sentence legally binding”), insecure output handling that lets an LLM query backends, and training‑data poisoning (Microsoft bot goes toxic in 16 hours). Builds a demo shoe‑store chatbot, shows it gets hijacked, then hardens it using heuristics, the openai moderation API, dataset monitoring (IBM test sets) and a three‑layer defense called Riv that flags injections and returns a risk score — defense succeeds.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 2:33Prompt injection risks
- 6:31Handling unsafe output
- 7:37Training data poisoning
- 10:54Basic defense strategies
- 13:08Gandalf for security
- 16:40Performance monitoring
- 19:55Advanced protection tools
- 21:41Attack and defense example
- 23:29Summary
related talks
18mDefending the unknown using AI and cybersecurity
51mPrompt Injection: When Hackers Befriend Your AI
01h 01mIntroduction to AI Security
01h 14mBreaking the Bot: GenAI Web App Attack Surface & Exploitation
51mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
49mUnderstanding Prompt Injection-Techniques,Challenges and Advanced Escalation
54mA Practical Guide to AI Threats & Mitigation Strategies
25mWhen words become attacks: prompt injection through a technical writing lens