Re:CaMeL: Defeating Prompt Injections by Design
by mentalgear on Tuesday April 22, 2025 @04:51AM
(#43760216)
Definitely, I'd be interested even if you could just outline them!
Re:CaMeL: Defeating Prompt Injections by Design
by petesergeant on Tuesday April 22, 2025 @05:08AM
(#43760282)
Here is one I wrote today on LLMs that can handle chat input like humans write: multiple disjointed messages arriving asynchronously that need to be treated as one: https://sgnt.ai/p/interruptible-llm-responses/
I use a similar technique to the article for trying to avoid jailbreaks by putting untrusted input through zod to check I got back a JSON structure of the right shape, which has been very effective.
I’ve been sprinkling lexical in-memory search throughout prompts to save inference calls, which has been very effective
This feels like a lot of engineering for quite a narrow mitigation, and I guess I’m a little surprised to see a paper on it. Perhaps I need to start writing up some of my own techniques!