Study probes what safety-aligned LLMs learn from mixed compliance demonstrations

arXiv·medium signal

Dai and Patel investigate how safety-aligned LLMs interpret in-context demonstrations that mix compliant and non-compliant behavior, building on prior work showing demonstrations can jailbreak models. The paper clarifies which demonstration types actually shift model behavior, sharpening understanding of in-context jailbreak risk for agents that consume untrusted context. It is relevant to any agent ingesting tool output or retrieved documents that could embed adversarial 'compliance' examples.