A new arXiv preprint describes a method for making text behaviorally invisible to language models. The authors show that when a piece of text is wrapped in a specific trigger, the model acts as if the text is not there: it is left out of summaries and quotations, ignored when answering questions, and not flagged by downstream filtering systems.
The paper calls this effect "selective blindness" and frames it as both covert and provable. The abstract does not detail the trigger mechanism or the conditions under which it works, but it states the phenomenon clearly across multiple tasks, suggesting a general vulnerability rather than a single model quirk.
Because this is based on a single source and only the abstract is available, the practical scope and reliability of the method remain unclear. Still, the claim that hidden text can evade both generation and filtering has implications for content moderation, data provenance, and model auditing.