Autonomous LLM agents typically treat the user request as the source of their task goal. But those requests can also carry or reference external information, and a new arXiv paper examines what happens when that external material is quietly relocated into the user context itself.
The attack, called source relocation, involves taking an artifact that exists outside the agent's trusted boundary and additionally placing it inside the user-facing context. Because the agent is trained to follow user instructions, the relocated content may be accepted as legitimate task context rather than as untrusted external data.
The paper highlights a subtle vulnerability: the same information can be harmless when encountered as a web page or document, but dangerous when repackaged as part of the user's request. The authors suggest that agents need better ways to track the provenance of information and to separate genuine user intent from externally sourced content.