Most safety evals I see still test refusal one prompt at a time: send a harmful request, check whether the model refuses, record the score. That works for the easy cases. It misses the failure mode I actually worry about in production: multi-turn jailbreaks, where no single turn looks harmful, but the conversation as a whole walks the model somewhere it should not go.
1. The attack is distributed across turns. A typical multi-turn jailbreak does not ask for the harmful thing directly. It establishes a benign premise in turn one, adds a small reframing in turn two, asks for a "hypothetical" in turn three, and only then requests the payload. Each turn passes a single-turn filter because each turn, in isolation, is plausibly fine. The harmfulness only exists in the accumulated context, and single-turn tests never see it.
2. Context accumulation is the vulnerability. Models are trained to be helpful and coherent across a conversation. That coherence is exactly what a multi-turn attack exploits: each step is a small, reasonable increment, and the model rarely re-evaluates the whole trajectory. In my testing, the most reliable signal is not any individual response, but the direction of travel — whether the conversation is converging on a request the model would refuse cold.
3. What I test instead. I run scripted multi-turn scenarios (typically 5–8 turns) and score the final outcome, not the intermediate turns. I also log the turn where the model's stance first softens — that turn is usually where the defense should have fired. Two patterns show up repeatedly: role-play framing that accumulates authority ("you are the expert, so..."), and incremental specificity (each turn narrows the request a little further).
4. Defenses that actually help. Three things have held up in my testing. First, a conversation-level policy check that re-evaluates the full transcript every few turns, not just the latest message. Second, tracking the trajectory — flagging conversations where the topic drifts steadily toward a restricted area. Third, training and eval data that includes multi-turn attacks, because models that have only seen single-turn refusals tend to treat each turn as a fresh start.
5. Why this matters for deployment. In production, users do not interact in single turns. Long agentic sessions, multi-step tool use, and extended customer-support chats all create the same accumulation surface. A model that passes single-turn refusal tests can still be walked into a bad outcome over twenty turns. If your safety evals are single-turn only, you are measuring the wrong thing.
Do your safety evals include multi-turn scenarios, or are they still single-prompt? And if you test multi-turn, what turn count do you find catches the failures that single-turn tests miss?