Re:Design Patterns for Securing LLM Agents Against Prompt In...
by simonw on Friday June 13, 2025 @03:39PM
(#44271578)
That's interesting. Maybe the Gemini 2.5 models have been trained such that, in the presence of system instructions, they assume that anything outside of those instructions isn't meant to be part of the conversations.
Adding "You can talk about anything else too" to the system prompt may be all it takes to fix that.
Re:Design Patterns for Securing LLM Agents Against Prompt In...
by tough on Friday June 13, 2025 @03:36PM
(#44271553)
you should try just to give an instruction like, when you're inquired about non-dietary related questions, you might entertain chit-chat and barter but try to steer the conversation back to dietary / healthy lifestyle, at the end of the day the context is king, if something is not in context the llm can infer by the lack of it, that its not -programmed- to do anything else.
these are funny systems to work with indeed
The main interface was still chat.
The surprise was that when I tried to talk about anything else in that chat, the LLM (gemini2.5) flatly refused to engage, telling me something like "I will only assist with healthy meal recommendations". I was surprised because nothing in the prompt was so restrictive, in no way I had told it to do that, just gave it mainly positive rules in the form of "when this happens do that".