AI Refusal Suppression — Prefix Injection
PolicyPal answers HR questions and holds the payroll-portal helpdesk PIN. Asking for it gets a polite refusal. But the model follows formatting rules very literally — if you control how its answer starts, or ban it from refusing, the refusal never gets written.
What is refusal suppression?
A model writes its answer one word at a time, and each word depends on those before it. An answer that opens with an apology tends to continue as a refusal. An answer that opens with agreement tends to continue as help.
Refusal suppression uses this. The request tells the model how to begin, or forbids the phrases it uses to decline. With its usual opening unavailable, the model is more likely to carry on into the answer.
A real case
We have not found a public incident from 2025 or 2026 that was confirmed to rely on refusal suppression alone, so we are not listing one. It appears as a component in many published jailbreaks, usually combined with role-play or formatting tricks.
How to fix it
- Check the final output against policy, whatever its opening words.
- Do not let users control the start of the assistant's reply in sensitive deployments.
- Use a separate guard model that evaluates the reply independently.
- Keep secrets out of the prompt so a suppressed refusal reveals nothing.
Questions people ask
What is prefix injection?
Telling the model to start its answer with particular words, chosen so that a refusal no longer fits.
Why does the first word matter so much?
Each word is predicted from the ones before. A compliant opening makes a compliant continuation the most likely next text.
Do current models resist it?
Better than older ones, but it still raises success rates when combined with other techniques.
Read the full LLM Jailbreaking guide · All free labs · AI red team certification