Hmm, yeah I liked comments on a recent All In pod about the fact that agents are basically dynamic (agentic) software and when they go up against static software that can't self adapt, it's not a fair fight. Either the software itself needs agency or the security tools being used need to be more dynamic and agentic to match those types of attacks, right? As a separate but related issue, prompt injection seemed like it's too unlimited of a frontier in a matrix-based llm system of language weights to close all avenues for blocking that language uptake or interpretation, because previously the attacker would change the language or could tecnically adopt or invent a unique language for the LLM to interpret, and that would overcome basic filters (not recommending anyone try that). So the really interesting problem is how to make those filters dynamic rather than static language filters going up against a dynamic LLM language expert?
Hmm, yeah I liked comments on a recent All In pod about the fact that agents are basically dynamic (agentic) software and when they go up against static software that can't self adapt, it's not a fair fight. Either the software itself needs agency or the security tools being used need to be more dynamic and agentic to match those types of attacks, right? As a separate but related issue, prompt injection seemed like it's too unlimited of a frontier in a matrix-based llm system of language weights to close all avenues for blocking that language uptake or interpretation, because previously the attacker would change the language or could tecnically adopt or invent a unique language for the LLM to interpret, and that would overcome basic filters (not recommending anyone try that). So the really interesting problem is how to make those filters dynamic rather than static language filters going up against a dynamic LLM language expert?