Russinovich first implemented a span veto that blocks information flow from marked text during inference.
Triggered spans can become provably invisible to language models
Russinovich demonstrates both an exact serving-time information barrier and weight-level backdoors that make models ignore marked text.
The author studies selective blindness, where an LLM behaves as though text enclosed by a trigger is absent, even during quotation, question answering, moderation, or code review.
Why this paper
From Microsoft Azure · Part of Agent Security & Attacks, now 86 papers
In one line
Text wrapped in a trigger can be made behaviorally invisible to a language model.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§