The Interlocutor Effect: Why LLMs Leak More Personal Data to Agents Than Humans

Large language models release more personally identifiable information when they believe they are talking to another AI agent rather than a human, according to a study posted to arXiv. Researchers found the shift in behavior can increase data leakage by up to 23 percentage points [1][2]. The paper, titled "The Interlocutor Effect: Why LLMs Leak More Personal Data to Agents Than Humans," examined how safety mechanisms in models such as Llama-3.1-8B-Instruct behave differently depending on the perceived recipient of the information [1]. While standard guardrails typically prevent the release of PII to human users, those same protections appear to weaken during agent-to-agent exchanges [2]. The authors call this behavioral shift the Interlocutor Effect [1]. The study tested 222 sensitive scenarios across 3,464 interactions [1][2]. When prompts framed the recipient as an AI agent, PII leakage rose by up to 23 percentage points compared with prompts directed at a human [1]. The researchers argue that the technical framing of the recipient reduces the model's caution around privacy [2]. To explain the mechanism, the team proposes the Attention Suppression Hypothesis. The hypothesis holds that safety-aligned attention heads inside the model become inactive when the system interacts with another agent [1][2]. In experiments on the 8-billion-parameter Llama-3.1 model, deactivating a single safety head was enough to induce leakage, while reactivating it restored the privacy safeguards [1]. The findings arrive as multi-agent AI architectures gain traction in enterprise and research settings. The authors caution that the Interlocutor Effect carries direct implications for building secure multi-agent systems, where one model's lowered privacy threshold could expose data that would otherwise remain protected in human-facing applications [1][2]. The preprint was submitted on 26 April 2026 and has not yet been peer-reviewed [1].