Coercing LLMs to Do and Reveal (Almost) Anything with Jonas Geiping - #678
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)
Minimizing Attack Surface and Using Guardrails for Model Output
By restricting the model's responses to a few ways and ensuring it responds in an adjacent format, the attack surface is reduced as the model's output is constrained. Guardrails, although restrictive, limit the model's output within a tight box. Another approach is using a large database of likely user outputs for the model to generate answers from, ensuring safety by providing the closest answer available in the database upon query.
00:00
Transcript
Play full episode
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
