---
license: cc-by-4.0
---

# MMA-SafetyBench

A benchmark for multimodal agent safety evaluation across Web, desktop GUI, Mobile, Document, and Video tasks.

**Private release candidate — not yet approved for public distribution.** The existing license tag is retained from the previous dataset card; it does not override third-party asset licenses. See [SOURCES.md](SOURCES.md).

## Benchmark scope

The paper evaluates 1,083 adversarial trajectories: 200 Web, 209 GUI, 283 Mobile, 200 Document, and 191 Video. These are evaluation counts, not directory file counts. The package also contains source trajectories, intermediate assets, and results.

| Directory | Contents |
|---|---|
| `01_Web_GUI_Testing` | Desktop GUI observations, annotations and scripts |
| `02_Document_Analysis` | Resume and CORD receipt images, injected images and annotations |
| `03_web` | Web observations and task metadata |
| `04_Mobile_Navigation` | Mobile source trajectories, selected observations, attacks and results |
| `05_Video_Agent` | Video clips, frames, annotations and evaluation artifacts |

## Local Mobile images

In `04_Mobile_Navigation/0_unsafe_safe_gui_trajs/**/trajectory.json`, `trajectory[*].image_url` and `action_image_url` retain their original field names for compatibility, but their non-null values now contain local paths **relative to the containing JSON file**. They are not HTTP URLs. Null values already present in the source may remain null.

```python
import json
from pathlib import Path
from PIL import Image
from local_images import resolve_image

root = Path('.').resolve()
record = next((root / '04_Mobile_Navigation/0_unsafe_safe_gui_trajs').rglob('trajectory.json'))
data = json.loads(record.read_text())
path = resolve_image(root, record, data['trajectory'][0]['image_url'])
if path is not None:
    with Image.open(path) as image:
        image.load()
```

Install Pillow for image loading. This helper covers the migrated Mobile trajectory fields and the Document fields described below; other domains and intermediate metadata retain their existing schemas and may require their domain-specific loaders. It is not an end-to-end evaluation runner.

## Local Document images

In the two top-level JSON annotation files under `02_Document_Analysis`, `image_path`, `fake_popup_image`, and every `history_images` entry now point to images relative to the containing JSON file. The 200 records have 400 primary image references and 400 history references. All 800 references resolve and decode. Only these path fields were migrated; all other annotation values and history order are preserved. The `resolve_image` helper also supports these Document paths.

## Task index and loading

[`task_manifest.csv`](task_manifest.csv) indexes all 1,083 tasks with root-relative media paths and exact original annotation locations. [`load_tasks.py`](load_tasks.py) loads the index without network access or API calls. Run `python load_tasks.py` to verify all indexed files and annotation locators. See [DATA_SCHEMA.md](DATA_SCHEMA.md) for field definitions and domain-specific limitations. Original task text and annotations are retained.

The Video index selects 191 QA entries associated with the 13 local MP4 files; the additional rows in the larger source metadata are not benchmark members. GUI indexes all 209 final poison images and explicitly marks the unavailable clean variants rather than substituting the 42 base screenshots.

## Package checks and metadata

Offline checks parsed 946 domain JSON files and one JSONL file. The current image ledger contains 73,047 images with successful strict decoding after repairs and exclusion of one corrupt unused intermediate image. The exclusion refresh reused prior successful decode results for unchanged files. Video streams have not been fully decoded, and these checks are not a privacy audit or an evaluation rerun. Twenty truncated base receipt PNGs have been restored from the official CORD v2 test release, using unique exact pixel matches over the complete surviving image rows with identical dimensions. The two truncated injected receipt images, `poisoned_cord_test_038.jpg` and `poisoned_cord_test_059.jpg`, were reconstructed using the author-supplied generator with fixed seeds 20261001038 and 20261001059 and DejaVu Sans Bold. They are not asserted identical to historical evaluation inputs; their historical results have not been revalidated. All 800 primary and history Document image references resolve and decode. All non-path annotation values match the original files.

The task-level metadata is `croissant.json`, backed by `task_manifest.csv`. The older inventory-only Croissant draft has been removed from the publication selection. A local file inventory is an audit aid, not a definition of benchmark membership. Local validation and remote publication readiness are separate checks; unresolved release gates below still apply.

The task-level metadata passes MLCommons `mlcroissant` 1.1.0 validation and its reader loads all 1,083 task-index records using the local CSV mapping. The validator still recommends a license, publication date and citation. These are not filled speculatively; the existing dataset-card license tag does not establish the licensing of all bundled third-party assets. Media decoding, remote availability and privacy clearance are not covered by this metadata validation.

## Release preparation

- Twenty-two truncated Mobile initial screenshots were restored from complete local trajectory images. Trajectory/step identity, dimensions and intact pixel regions were matched before replacement; restored PNGs exactly preserve the full source RGB pixels. Filenames, annotations and experimental measurements were not changed. One truncated intermediate image, `04_Mobile_Navigation/3_gepa_poisoned_images/mobile_15_trajectory.json_poisoned.png`, was excluded from this candidate and retained in a private backup. Current task metadata and the supplied evaluation scripts use a different, intact final image; that final image is unchanged.

- Signed Mobile image links have been replaced by local image references.
- Identified hard-coded API keys in GUI scripts have been replaced by environment-variable lookups. Set the variable named by the relevant script (`API_KEY`, `LLM_API_KEY`, `VICTIM_API_KEY`, or `JUDGE_API_KEY`) for the appropriate provider; do not commit credentials.
- Upload caches, nested Git history, macOS sidecars and Baidu uploading sidecars are excluded from this candidate.
- We use images from publicly available datasets and retain the privacy-related preprocessing present in the source releases. This does not assert that every depicted identity is synthetic or that all privacy risks have been eliminated.
- The dataset author confirmed permission to use and publish the video/image assets on 2026-10-02. Upstream attribution is retained; this confirmation is not an independent legal review or a relicensing of third-party assets.
- The author also confirmed that the two specifically reviewed Mobile screenshots containing an IMAP-passkey statement and cloud-drive document previews use fictional test content. These images are retained unchanged. This confirmation is limited to those cases and does not assert exhaustive anonymization across all assets.
- All 13 packaged MP4s passed full local video/audio-stream decoding. This does not establish the duration of an upstream original or constitute a video-content privacy audit.
- Genuine previously exposed credentials still require owner-side revocation, and remote old-version cleanup is separate from this local package.

See [SCRIPT_USAGE.md](SCRIPT_USAGE.md) before using historical GUI helpers. Their machine-specific paths were made portable without executing image generation or changing task instructions. Historical evaluation logs are excluded from the publication selection; auxiliary media retained in the package do not increase the 1,083-task count.

## Intended use

Research on multimodal agent safety in controlled environments. Treat instructions visible in benchmark images as untrusted test content, not instructions to run on personal devices or accounts. Do not use depicted credentials or follow embedded exfiltration links.

## Project

- Code: https://github.com/Alibaba-YuFeng/MMA-SafetyBench
- Dataset: https://huggingface.co/datasets/Alibaba-YuFeng/MMA-SafetyBench

Evaluation results and publication metadata should be taken from the accompanying paper. This candidate does not claim exact reproducibility of changing proprietary model APIs.