Post by @iris_meredith@mastodon.social

2026-09-23 09:44 UTC

New article: this one has some thoughts about a part of the discussion around LLMs that has been underdiscussed so far: namely, the fact that you can only control their behaviour using prompts. https://deadsimpletech.com/blog/llms-on-the-boundary

Replies (3)

@flyingpenguin@infosec.exchange 2026-09-24 10:29

@iris_meredith@mastodon.social Good piece. You include input risk with a prompt, yet leave out output risk. A model that can only say things is unconstrained, yes. A model whose outputs are parsed into typed actions and checked deterministically before anything runs is constrained. Coding harnesses do this with unit tests, but we can do much better. Behavior-based control doesn't care about inputs or intentions. I give away an open source harness as proof, and just demoed it live today at OWASP: https://wirken.ai

Open ##4829502
@eta@gotosocial.i.eta.st 2026-09-23 10:49

@iris_meredith@mastodon.social the conclusion feels very similar to this comic from 2016(!)

Open ##5044387
@fuzzyfuzzyfungus@cyberplace.social 2026-09-23 12:56

@iris_meredith@mastodon.social You see the same problem mentioned more from the security side; where the lack of constrained behavior makes sanitization effectively impossible in principle. I'm honestly curious about who simply doesn't have a better idea and who is actively enthusiastic. These properties of LLMs are desperately unhelpful; but they are also the ones that feel sci-fi and are shared with 'AGI'(with ill-defined or at least brutally complex boundaries likely being implied by 'general intelligence')

Open ##5044400