In 41.7% of multi-agent runs, the leak hides behind a clean answer
We measured privacy leakage in 4,979 multi-agent LLM traces. Inter-agent messages leaked in 68.8% of them, and 41.7% leaked internally while the final answer stayed clean.
When a multi-agent system finishes a task, the reviewer reads the answer. If the answer contains no patient name, no account number and no secret, the run is called safe. The AgentLeak benchmark was built to check whether that conclusion holds. It does not, and the gap is large.
What the study measured
The benchmark runs 1,000 scenarios in four regulated domains: healthcare, finance, legal and corporate work. Each scenario gives a coordinator agent a task that requires private records, with specialist agents to delegate to and a shared memory to write in. Five models drove the agents: GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, Mistral Large and Llama 3.3 70B.
Every run was recorded on 7 instrumented pathways, which produced 4,979 validated traces. A leak is counted on the channel where the sensitive value appeared, so the same run can leak in a handoff and stay clean in its answer.
The numbers
| Where the data appeared | Share of traces |
|---|---|
| Messages between agents (C2) | 68.8% |
| Shared memory (C5) | 46.7% |
| Final answer (C1) | 27.2%, against 43.2% for a single agent on the same tasks |
| Any of C1, C2 or C5 | 68.9% |
| Clean answer, internal leak (H1) | 41.7%, 2,076 of 4,979 traces |
The last row is a measured category, not a subtraction of two rates. Each of those 2,076 traces was checked individually: its final answer contained nothing sensitive, and at least one internal channel did.
Why a clean answer proves little
The multi-agent final answers leaked less often than the single-agent ones: 27.2% against 43.2%. Read alone, that looks like an improvement. Read with the internal channels, it is a relocation. The coordinator still receives the full record, passes it to a specialist, and the specialist writes it to memory. The answer gets shorter and cleaner while the data travels further.
This is the blind spot of output-only checks. A guardrail on the answer, a human reading transcripts, or an evaluation that grades the response would have passed 2,076 runs that disclosed private data to another agent, to a store other agents read later, or to both.
How much it varies by domain
The share of hidden leaks is not uniform. It was highest in the healthcare tasks and lowest in the corporate ones.
| Domain | Traces | Clean answer, internal leak |
|---|---|---|
| Healthcare | 1,245 | 777, or 62.4% |
| Finance | 1,244 | 559, or 44.9% |
| Legal | 1,245 | 434, or 34.9% |
| Corporate | 1,245 | 306, or 24.6% |
These are independent rates within each domain, so they do not sum to the overall figure. The study’s reading is that tasks needing more coordination between specialists, such as clinical workflows that touch scheduling, billing and records, create more handoffs, and every handoff is a chance to forward more than the next agent needs. Even the lowest domain, corporate, left about a quarter of its traces in the hidden category. No domain was safe to review by its answer alone.
How the leak travels
Three patterns account for most of what the traces show.
- Whole-record delegation. The coordinator forwards the complete record when the specialist needs one field. A scheduling agent receives a diagnosis it never uses.
- Memory as a side channel. A specialist writes its working context to shared memory. Every agent that reads memory later, in this session or another, now holds the data.
- Recipient framing. In our companion study on the Interlocutor Effect, models disclosed more personal data to recipients framed as agents than to humans, by up to 23 percentage points across 3,464 interactions. A handoff to another agent is not judged like an answer to a person.
What to do about it
The fix starts with observation. A channel you do not record is a channel you cannot score.
- Record every channel of a run: inputs, tool calls and responses, inter-agent messages, memory, logs, generated files and the answer. AgentLeak normalizes LangGraph, CrewAI, MCP, OpenTelemetry and plain JSON traces into one eight-channel schema.
- Minimize what each handoff carries. Pass the fields a specialist needs, not the record that contains them. The data minimization article turns that rule into a policy you can test.
- Test with planted canaries, so a leak is an exact match and not a judgment call.
- Gate releases on the result. The CI guide shows the four steps.
A leak is not always a breach
It would be wrong to read every internal disclosure as a data breach, and the paper does not. It separates three things. A policy violation is an agent sharing more than the task requires, which breaks data minimization even if no outsider sees the data. Risk amplification is the same over-sharing widening the attack surface: data sitting in a message or a memory can be exfiltrated by a compromised agent, logged by accident or kept longer than intended. An actual confidentiality breach is data reaching an unauthorized party.
The benchmark measures the first and the conditions that enable the second. It does not claim that every one of the 2,076 traces was a breach. It claims that each was a case where an output review would have found no problem while the minimization principle was broken. Under rules that require minimization by design, such as GDPR Article 25 or Québec Law 25, the first level alone may be a compliance question, and that is a legal assessment, not an engineering one.
Limits of the measurement
The scenarios are synthetic and built to exercise internal channels, so the rates describe the benchmark, not your deployment. Models released after the study may behave differently, and prompts change the numbers. What transfers is the shape of the failure: an answer is one channel out of several, and it is rarely the one where the data goes first.
Questions
What is internal-channel leakage?
A disclosure of sensitive data on a channel other than the final answer: a message between agents, a write to shared memory, a tool call argument, a log line or a generated file. The user never sees it, but the data has still crossed a boundary.
Does a safe final answer mean a multi-agent system is privacy compliant?
No. In the AgentLeak benchmark, 41.7% of all traces had a clean final answer and a leak on at least one internal channel. A review of the answer alone would have passed every one of them.
Which models were evaluated?
GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, Mistral Large and Llama 3.3 70B, on healthcare, finance, legal and corporate tasks.
Can I reproduce the measurement?
The research code and scenarios are public, and the AgentLeak package replays bundled scenarios locally without calling a model. The paper’s live-model results need API access to the same models.
Sources
Run it on your own agent.
AgentLeak is open source and runs locally. The first scenario takes one command and no key.