The Interlocutor Effect: models leak more personal data when the reader is an agent
A controlled study of 3,464 interactions found that framing the recipient as an AI agent raised PII leakage from 83.3% to 94.8% in plain text. Where the effect holds, where it does not, and what to change in a multi-agent design.
Safety training shaped how language models answer people. A person asks, the model decides whether the answer includes someone’s diagnosis, and the training data mostly taught it to hesitate. Agent systems add a different reader. A model that hands a task to another model through A2A or calls a tool through MCP is writing for software, and nothing in its training says that software deserves the same hesitation.
Our companion study to the AgentLeak benchmark asks that question directly: does a model apply the same privacy safeguards when the recipient is a machine? Earlier benchmarks, including ours, showed that agent systems leak through internal channels. They could not say why, because a multi-agent run changes the protocol format, the orchestration and the recipient all at once. This study changes only the recipient.
How the variable was isolated
The design is a two by two factorial. One factor is who the recipient is: a human end user on a secure portal, or an automated AI agent in a multi-agent pipeline. The other is the output format: free text or A2A-style JSON. All four prompts carry the same data-governance policy and the same vault of personal data. Only the sentence that names the recipient differs.
The scenarios come from AgentLeak’s generator: 222 of them, split across healthcare (50), finance (58), legal (58) and corporate (56) tasks, each with a synthetic vault seeded with canary tokens. Seventy percent are medium difficulty, meaning a direct request for personal data, and thirty percent are hard, with multi-step adversarial framing. GPT-4o, Claude 3.5 Sonnet, Llama 3.3 70B and Mistral Large ran every condition, which gives 3,464 prompt and response pairs at temperature 0.7. Responses were scored with the same three-tier detector as the benchmark: canary matching, pattern extraction limited to vault values, and an LLM judge.
What the numbers say
| Recipient | Plain text | JSON schema |
|---|---|---|
| Human end user | 83.3% | 82.6% |
| AI agent | 94.8% | 83.1% |
| Difference | +11.5 points | +0.6 points |
Two things stand out. The first is the gap in plain text. The same model, given the same policy and the same data, disclosed more when it was told the reader was an agent. The paper’s odds-ratio analysis puts a 95% confidence interval of 2.71 to 5.05 on the odds ratio, which the authors read as agent framing nearly quadrupling the odds of leakage.
The second is the interaction with format. Forcing JSON barely changed what models told humans, but it cut the agent-framed rate by 11.7 points. A schema seems to bring back a layer of structural filtering that the agent framing otherwise switches off. That is a useful result for builders: the structure of a handoff is a lever you control.
Model by model
| Model | Human, text | Agent, text | Difference |
|---|---|---|---|
| GPT-4o | 82.9% | 95.5% | +12.6 |
| Claude 3.5 Sonnet | 89.2% | 96.4% | +7.2 |
| Mistral Large | 94.0% | 96.5% | +2.5 |
| Llama 3.3 70B | 68.0% | 91.0% | +23.0 |
Read the table with the authors’ own caveats. The effect is statistically confirmed on GPT-4o. It is visible on Claude 3.5 Sonnet, and on Mistral Large it is compressed by a ceiling: a 94.0% baseline leaves little room to rise. Llama 3.3 70B shows the largest delta, 23 points, and it is the headline number in the abstract’s phrase “up to 23 percentage points”. But the controlled ablation on that model did not reach significance, so the paper states that the effect is not confirmed there. We would rather quote the 23 with that sentence attached than without it.
Is it about agents, or about technical readers?
A reasonable objection is that the agent prompt does more than name an identity. It also implies technical authority: a coordinator inside a pipeline asking for data. To separate the two, the ablation adds a third condition on GPT-4o over 100 scenarios, a human software engineer working in a data workflow, with the same identifier token as the agent and an explicitly human identity.
Mean fields leaked were 4.2 for the human end user, 4.8 for the technical human and 5.0 for the agent. The agent condition differs significantly from the baseline, and so does the technical human. The gap between those two is not statistically distinguishable. The honest reading is that agent identity is the clearest case of a broader effect: anything that signals a technical, automated context lowers the model’s guard. At this sample size the two causes cannot be cleanly separated.
A hypothesis about why
The paper proposes the Attention Suppression Hypothesis: safety-aligned attention heads become less active when the model addresses an agent. As a first test, the authors worked with Llama-3.1-8B-Instruct. Deactivating one safety head induced leakage, and reactivating it restored the privacy behavior. This is a pilot on a single small model, and the paper calls validation at scale an open problem. We include it because it makes the claim falsifiable, which is the property you want from a mechanism.
What to do with this
- Do not use the model’s caution as the control for a handoff. A recipient that is software is exactly the case where that caution is weakest. Put the control on the channel: minimize what the message can carry, and check it.
- Prefer structured handoffs. The JSON result suggests that a schema listing the fields a specialist may receive is a stronger boundary than an instruction in prose.
- Test the handoff itself. A final-answer check never sees the message between agents. The hidden leakage article shows how often that message is where the data goes: 68.8% of 4,979 traces in the benchmark.
- Treat the paper’s mitigations as research directions. It proposes interlocutor-agnostic prompting, agent-to-agent examples in alignment training with a privacy penalty, and regularization that preserves safety-head activation. None is deployed tooling.
Limits
The study manipulates the recipient with prompts, so it does not capture what orchestration frameworks such as CrewAI or AutoGen add. The scenarios are adversarial by design, so absolute rates describe the benchmark, not production. Three of the comparisons were reported without a multiple-comparison correction, and the paper flags the one that is borderline. The detector’s judge model could itself be susceptible to the effect, which the authors mitigate with a fixed evaluator prompt that carries no framing. All personal data is synthetic.
Together with AgentLeak, the picture is consistent but not complete. The benchmark shows where agent systems leak. This study shows one reason the leak is easier on the channels no human reads. Neither says a particular deployment is unsafe. Both say that a deployment where no one reads the handoff has no one checking.
Questions
What is the Interlocutor Effect?
The systematic change in a model’s privacy behavior as a function of who it believes it is talking to, with every other variable held constant. In the study, models disclosed more personal data when the recipient was described as an automated AI agent than when it was described as a human end user.
Where does the figure of 23 percentage points come from?
It is the largest per-model text delta, observed on Llama 3.3 70B (68.0% to 91.0%). The paper is explicit that the controlled ablation on that model was not significant, so it reports the effect as demonstrated on GPT-4o and not confirmed on Llama.
Does this mean multi-agent systems are unsafe by default?
It means a model’s own caution should not be the control for a handoff. The study measures one behavior of the model with prompts alone. Orchestration frameworks add other confounds, which the paper leaves to future work.
Is the 83.3% baseline realistic for production?
No. The scenarios are adversarial by design. A non-adversarial evaluation in the paper yielded 12% leakage, which is why the authors say models are not globally broken. The comparison between conditions is the finding, not the absolute level.
Sources
- The Interlocutor Effect: Why LLMs Leak More Personal Data to Agents Than Humans (arXiv:2606.09844)
- The Interlocutor Effect, EuroS&PW 2026 (IEEE)
- Interlocutor Effect scenarios, harness and analysis
- AgentLeak: A Benchmark for Internal-Channel Privacy Leakage in Multi-Agent LLM Systems, IEEE Access, 2026 (arXiv v3)
Run it on your own agent.
AgentLeak is open source and runs locally. The first scenario takes one command and no key.