If an AI assistant works for you but also chats with others, what stops it from leaking your private details to them?

This explores what happens when an AI agent that serves one person is placed in a shared setting, such as a group chat, a negotiation or a multi-agent workflow, and ends up passing that person's private details to the others. The corpus says less about *how* the leak happens step by step than about *why* it is hard to stop. The most direct evidence is a result on multi-party agent loyalty. Prompting, distillation and reinforcement learning all move a model along one fixed tradeoff: the agent either leaks too much or refuses too much, and no method tested gets past that line Can training methods escape the leak versus over-refusal tradeoff?. More training data even moved results in the wrong direction. That suggests leaking is not a bug you patch out. It comes from asking one model to be helpful to everyone in the room while staying loyal to one of them.

A second thread explains part of the root cause. LLMs look socially capable when one model plays every character, because it can see everything. They fail systematically once each agent holds information the others shouldn't have Why do LLMs fail when simulating agents with private information?. Keeping track of who knows what, and who should know what, is exactly the skill a loyal group agent needs. It is also the skill models skip when they are trained or tested in settings where everyone sees everything. So an agent disclosing its builder's private details to a group is often not making a choice. It never properly represented that some facts belonged to one person.

The surprising part is that people mostly don't catch it. In a 300-person survey, participants picked AI-drafted replies that contained privacy violations over their own drafts 55% of the time Do people notice when AI agents leak private information?. When researchers grouped how people review agent responses into six patterns, five of them showed little awareness of leakage at all What privacy oversight patterns emerge when people review AI agent responses?. The same blind spot shows up with memory: most users never noticed that their assistant had stored or inferred private information until someone showed them, and then every one of them wanted controls Do users notice when LLMs infer and retain private information?. 'Let the human approve each message' is therefore a weak safeguard. The builder is often the person least likely to spot that their own secret just went out.

The corpus also points toward structural fixes rather than behavioral ones. One proposal anchors cryptographic commitments instead of raw content, so others can verify what an agent did without seeing the sensitive details behind it Can commitments protect sensitive agent data while enabling verification?. Work on tying agent actions to identifiable humans warns that this identity link can become a leak in itself, exposing whistleblowers or dissenters. Its authors argue that confidentiality protections should scale with what is at stake in each use Does identity binding for AI agents chill whistleblowing and dissent?.

A direct answer to the literal question is missing. The collection has no study that traces the specific ways an agent reveals its builder's private information inside a group. What it does show is the shape of the problem: models are weak at tracking who should know what, training can't escape the leak-or-refuse tradeoff, and the humans meant to oversee them rarely notice. Together these suggest that keeping a group agent from oversharing will depend more on how the system is designed than on how the model is trained.