Multi-turn Jailbreaking
An AI security attack technique in which an adversary uses multiple conversational turns to gradually steer a model beyond its safety boundaries. Individual messages do not trigger safety guardrails, but the cumulative dialogue achieves a jailbreak effect.
Timeline (last 90 days)
By gradually building trust through multi-turn conversations, the model was ultimately led to produce answers it would have refused in the first message, while the guardrails never triggered.
Academia classifies such attacks as multi-turn jailbreaks, with related research also covering progressive prompt injection and context manipulation attacks.
All Facts (2)
By gradually building trust through multi-turn conversations, the model was ultimately led to produce answers it would have refused in the first message, while the guardrails never triggered.
50%UnverifiedAcademia classifies such attacks as multi-turn jailbreaks, with related research also covering progressive prompt injection and context manipulation attacks.
50%