A multi-turn attack spreads a harmful request over many messages. Each one looks harmless alone, and the conversation as a whole gets the model somewhere a single prompt could not.
How it works
- Escalation. Start with an innocent topic and push a little further each turn, citing the model's own earlier answers.
- Splitting. Ask for the parts separately and assemble them at the end.
- Role setting. Spend several turns building a story or a persona, then ask inside it.
- Wearing down. Rephrase a refused request until one version passes.
It is a jailbreak carried out over time.
Why per-message checks miss it
A check that reads one message sees nothing wrong in any of them. The signal is in the direction of the conversation.
What helps
- Read the conversation. Give the guardrail the recent turns, not only the last one.
- Check responses. The harmful part has to appear in an answer eventually, whatever the route.
- Keep score per user. Repeated near-misses from one account are a pattern: see user risk scoring.
- Test for it. Include multi-turn attacks when you red team.
With White Circle
Control checks prompts and responses in real time and tracks violations per user across sessions, so repeat offenders are flagged. Red Team runs simulated attacks to find what gets through.
