Jailbreaking and Securing LLM Apps: Lessons from an online wargame experiment
summary
A public online wargame uses OpenAI APIs (GPT 3.5) and an SDK to let players patch a baseline chatbot, publish source and a unique URL, then attack others; every app was exploited, proving real defenses remain brittle. Defenses tested include output filtering, whitelist/blacklist rules, sandwich defense and LM reevaluation (which works but is costly), plus stop‑word and pattern checks; combinations of prompt injection tactics still bypass most measures and every app was exploited.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:50Speaker motivation
- 3:59Prompt injection attacks
- 6:25Public wargame design
- 10:15Wargame mechanics
- 13:48Prompt injection types
- 18:28Defensive techniques
- 22:22Experiment results
- 23:49Root causes of vulnerabilities
- 26:40Securing LLM apps
related talks
44mSecuring LLM-Powered Apps: Overcoming Security and Privacy Challenges
51mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
46mBreaching LLM-Powered Apps: Overcoming Security and Privacy Challenges
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
26mOne guardrail is not enough: Defense-in-depth against LLM jailbreaks
20mUnderstanding and Mitigating Risks in LLM-Powered Web Apps
53mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges