Coercing LLMs to Do and Reveal (Almost) Anything with Jonas Geiping - 678

Today we're joined by Jonas Geiping, a research group leader at the ELLIS Institute, to explore his paper: "Coercing LLMs to Do and Reveal (Almost) Anything". Jonas explains how neural networks can be exploited, highlighting the risk of deploying LLM agents that interact with the real world. We discuss the role of open models in enabling security research, the challenges of optimizing over certain constraints, and the ongoing difficulties in achieving robustness in neural networks. Finally, we delve into the future of AI security, and the need for a better approach to mitigate the risks posed by optimized adversarial attacks.
🔔 Subscribe to our channel for more great content just like this: ua-cam.com/users/twimlai?sub_confi...
🗣️ CONNECT WITH US!
===============================
Subscribe to the TWIML AI Podcast: twimlai.com/podcast/twimlai/
Join our Slack Community: twimlai.com/community/
Subscribe to our newsletter: twimlai.com/newsletter/
Want to get in touch? Send us a message: twimlai.com/contact/
📖 CHAPTERS
===============================
00:00 - Introduction
02:20 - Are we ready for agents?
04:06 - Security and open-weight models
07:53 - How to make an LLM say anything
13:30 - What are the limitations?
16:06 - The role of code in vulnerability
18:17 - Interesting findings
22:41 - Prompt optimization
29:35 - RLHF and its possible alternatives
41:45 - Real-world impact of LLM vulnerabilities
46:40 - Where is this all going?
50:08 - Conclusion
🔗 LINKS & RESOURCES
===============================
Coercing LLMs to do and reveal (almost) anything - arxiv.org/abs/2402.14020
Gandalf (Jailbreaking game) - gandalf.lakera.ai/
World_sim - worldsim.nousresearch.com/
Mental Models for Advanced ChatGPT Prompting with Riley Goodside - 652 - twimlai.com/podcast/twimlai/m...
📸 Camera: amzn.to/3TQ3zsg
🎙️Microphone: amzn.to/3t5zXeV
🚦Lights: amzn.to/3TQlX49
🎛️ Audio Interface: amzn.to/3TVFAIq
🎚️ Stream Deck: amzn.to/3zzm7F5

[@stl8k]

Great point about tech-agnostic security considerations of freeform text input at 22 minute mark.

[@jdown330]

is bro talking on 1.25 speed? Wth