Doors of (AI)pportunity: The Front and Backdoors of LLMs
summary
Shows concrete exploits and defenses: prompt injection chains that stitch cultural references into SQL and image jailbreaks, an automated finder called autocrop that harvests those gadgets, a universal transferable alignment bypass named policy puppetry that overrides instruction hierarchies, and computational-graph backdoors that swap model behavior (e.g., Intellisense/np.load -> allow_pickle=True) or flip safety refusals. Demonstrations include a working RAG/Gemini workspace poisoning, reverse‑shell recovery from an image prompt, and minute‑scale local LLM jailbreaks on CPU.
AI-written summary · about
chapters (15)
- 0:00Introduction to AI doors
- 1:50How LLMs work
- 2:54Understanding prompt injection
- 5:26System prompts and defenses
- 7:24Techniques for model exploitation
- 9:28Advanced injection techniques
- 12:40Automating exploits
- 14:51Policy puppetry
- 18:20Indirect prompt injection
- 23:32Local model security
- 30:18Computational graph backdoors
- 34:54Inspecting computational graphs
- 38:01Backdooring computer vision
- 47:18Advanced LLM backdoors
- 52:07Takeaways and Q&A
related talks
49mDoors of (AI)pportunity: The Front and Backdoors of LLMs
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
39mLLMs everywhere: The future is now and the past keeps repeating itself
09mLLM hacking is underrated
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
36mCompromising LLMs: The Advent of AI Malware
27mSecuring The Gateway And Mitigating Risks In LLM API Integration
49mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges