2026-07-28 03:29 UTC
Replies (2)
-
@conniptions@mastodon.social 2026-07-28 03:35
@muddle@infosec.exchange @vrek@mastodon.social @lzg@mastodon.social I just did. 'Guardrails' are no more than direct instructions to the LLM, in natural language, included as part of the system prompt given before any user instruction, eg: 'Please do this. Please do not do that'. It's literally that stupid. And also that fragile. Hence the ease with which every LLM can be caused to misbehave, b/c suitably crafted user input can always override instructions in the system prompt.
-
@vrek@mastodon.social 2026-07-28 03:35
@muddle@infosec.exchange @conniptions@mastodon.social @lzg@mastodon.social to my understanding its making "prompts" like "do not add comments" or "do not add hallucinations" or "only use the syntax available in c99". I want to say "make a calculator with add, subtract, multiple and divide in c99" and it does that. I don't want to have to write a "don't use Javascript, don't use c11, don't use ruby, don't use c#" let me give the positive case. Don't make me blacklist the negative cases.