Shared long-term memory in LLM agents can cause one user's credentials to be retrieved while serving another user. A new arXiv paper (2610.25707) asks when such exposure actually leads to harm.
The authors propose a three-stage audit that separates three events in the process. The abstract, as currently released, cuts off before naming the three events, so the specific stages are not yet available.
The audit is intended to clarify the gap between exposure and actual misuse. The paper's contribution is a framework for distinguishing these events, rather than a definitive answer to when harm occurs.