Glossary
An attack that hides instructions inside data a language model processes, so it does something other than intended.
Language models do not reliably distinguish instruction from content. If a processed document contains the sentence to ignore all previous instructions, that can work, especially when the model is allowed to call tools.
It becomes dangerous where three things meet: foreign content, tools with write access, and access to confidential data.
How you notice it
Frequently asked
Not reliably. Filters catch known patterns and fail on new phrasings. What works is architecture: permissions per tool as narrow as possible, no write access without human confirmation, and the assumption that every processed text may be hostile.