2026-10-08 23:26 UTC
@Epic_Null@infosec.exchange @evacide@hachyderm.io that’s never gonna work. LLMs have no separation between instructions - what used to be called “code” - and “data” - the stuff it consumes while fulfilling your request. Even if their creators wanted to stop their bots behaving badly, someone’s going to figure out a way around it.
If you want a secure system, start with a secure system. Silicone Valley started with an academic toy. And think the taxpayer should bail them out for their mistake.
Replies (1)
-
@Epic_Null@infosec.exchange 2026-10-08 23:28
@Kynx@fosstodon.org @evacide@hachyderm.io LLMs have no separation between instructions - what used to be called “code” - and “data” - the stuff it consumes So don't put the triggers at the LLM level. Scan input for the codes before they hit the LLM. If the codes show up, take the relevant action.