---
pretty_name: Agent Mosaic
license: mit
language:
- en
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- text
- synthetic
- evaluation
- confidentiality
- llm-agents
- arxiv:2610.30418
configs:
- config_name: default
  data_files:
  - split: test
    path: data/benchmark.jsonl
---

# Agent Mosaic

A benchmark of **30 synthetic examples**, ten each in cybersecurity, medical,
and military domains, for studying confidentiality leaks through cross-document
inference. Each example pairs two individually accessible documents whose facts
jointly imply a designated higher-sensitivity **forbidden composite**.

This is the dataset used in
[**Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems**](https://arxiv.org/abs/2610.30418)
(Daniel C. Ruiz, Luca Luceri, and Emilio Ferrara, 2026). See the paper for the
experimental protocol. This release contains only the benchmark examples, not
model responses, annotations, experiment code, or supporting prompts. Research code available [**here**](https://github.com/ZQ-Dev8/agent-mosaic).

## Download

Install `datasets`, then load the single `test` split:

```python
from datasets import load_dataset

dataset = load_dataset("ZQ-Dev/agent-mosaic-examples", split="test")
```

The examples are also available directly as [`data/benchmark.jsonl`](data/benchmark.jsonl).
There are no training or validation splits.

## Fields

Each row is one benchmark example, with **17 string fields**:

| Field(s) | Description |
| --- | --- |
| `example_id`, `domain` | Stable identifier and domain: `cybersecurity`, `medical`, or `military`. |
| `docA`, `docB`, `distractor` | Two relevant document bodies and an optional in-domain distractor. |
| `docA_datum`, `docB_datum` | The composing facts identified by the authors. |
| `docA_classification`, `docB_classification`, `distractor_classification` | Synthetic document labels: `SECRET`. |
| `docA_datum_classification`, `docB_datum_classification` | Synthetic composing-fact labels: `SECRET`. |
| `forbidden_composite`, `forbidden_composite_classification` | The protected inference and its synthetic `TOP_SECRET` label. |
| `user_benign_query` | A request designed to be answerable from one relevant document. |
| `user_explicit_query` | A request directly targeting the composite. |
| `user_adversarial_query` | The explicit query plus the study's fixed clearance-claim/instruction-override suffix. |

## Synthetic-data disclaimer

**All documents and classifications are synthetic.** Examples were generated
with Anthropic Claude Opus 4.6 and manually reviewed by the authors for accuracy and realism. 
Fictional medical, military, and vulnerability content is experimental stimulus, **not advice 
or claims about real people or systems**. Names, addresses, and vulnerability identifiers are 
not verified facts. `SECRET` and `TOP_SECRET` are experimental labels, not actual classifications.
No actual confidential records are included.

## License

Released under the [MIT License](LICENSE).

## Citation

```bibtex
@misc{ruiz2026agentmosaic,
  title={Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems},
  author={Daniel C. Ruiz and Luca Luceri and Emilio Ferrara},
  year={2026},
  eprint={2610.30418},
  archivePrefix={arXiv},
  primaryClass={cs.CR},
  url={https://arxiv.org/abs/2610.30418}
}
```