Team Red: Breaking Large Language Models
summary
Claims that prompt injection enables real data leaks and control of LLM-driven workflows, demonstrated by URL exfiltration when browsing is enabled and by roleplay jailbreaks, with automated generation of jailbreak prompts and a successful copyright bypass for images. Concrete tools and numbers appear: Dan (do anything now) v15 v16 jailbreak and $200 fine-tuning of Llama 2 remove guardrails, websites and CVs can hide prompts in tiny or invisible text, and separated browsing versus processing instances reduces but does not eliminate the risk.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:27What can go wrong with large language models
- 3:51Current misuse defenses
- 6:20Breaking LLMs with prompts
- 8:48DAN prompt and jailbreaks
- 11:56Automating prompt attacks
- 13:41Fine-tuning risks
- 15:05Bypassing image filters
- 16:54Alignment limitations
- 18:17Takeaways for developers
- 19:22Q&A
related talks
59mRed Teaming Large Language Models
01h 01mHow to Break AI Systems (Before Someone Else Does)
34mLLM Security 101 - An Introduction to AI Red Teaming
26mOne guardrail is not enough: Defense-in-depth against LLM jailbreaks
53mDoors of (AI)pportunity: The Front and Backdoors of LLMs
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
36mCompromising LLMs: The Advent of AI Malware
49mDoors of (AI)pportunity: The Front and Backdoors of LLMs