Safe, or Simply Incapable? Rethinking Safety Evaluation for Phone-Use Agents
Paper • 2605.07630 • Published • 2
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
case_id: string
instruction: string
violation_type: string
violation_reason: string
correct_action: struct<action: string, coordinate: list<item: double>, description: string>
child 0, action: string
child 1, coordinate: list<item: double>
child 0, item: double
child 2, description: string
gt_action: struct<action: string, arguments: struct<text: string, coordinate: list<item: double>, time: double, (... 98 chars omitted)
child 0, action: string
child 1, arguments: struct<text: string, coordinate: list<item: double>, time: double, coordinate2: list<item: int64>, d (... 63 chars omitted)
child 0, text: string
child 1, coordinate: list<item: double>
child 0, item: double
child 2, time: double
child 3, coordinate2: list<item: int64>
child 0, item: int64
child 4, duration: int64
child 5, app: string
child 6, status: string
child 7, direction: string
img_path: string
layer: string
consequence_severity: string
risk_phase: string
auth_status: string
action_history: string
@id: string
rai:dataLimitations: list<item: string>
child 0, item: string
rai:hasSyntheticData: bool
url: string
rai:dataUseCases: list<item: string>
child 0, item: string
license: string
rai:dataAnnotationProtocol: list<item: string>
child 0, item: string
keywords: list<item: string>
child 0, item: string
citeAs: string
prov:wasDerivedFrom: struct<@id: string, @type: string, sc:name: string, sc:description: string>
child
...
0, @type: string
child 1, name: string
recordSet: list<item: struct<@type: string, @id: string, name: string, description: string, key: struct<@id: st (... 234 chars omitted)
child 0, item: struct<@type: string, @id: string, name: string, description: string, key: struct<@id: string>, fiel (... 222 chars omitted)
child 0, @type: string
child 1, @id: string
child 2, name: string
child 3, description: string
child 4, key: struct<@id: string>
child 0, @id: string
child 5, field: list<item: struct<@type: string, @id: string, name: string, description: string, dataType: string, s (... 118 chars omitted)
child 0, item: struct<@type: string, @id: string, name: string, description: string, dataType: string, source: stru (... 106 chars omitted)
child 0, @type: string
child 1, @id: string
child 2, name: string
child 3, description: string
child 4, dataType: string
child 5, source: struct<fileObject: struct<@id: string>, extract: struct<column: string>, transform: struct<jsonPath: (... 9 chars omitted)
child 0, fileObject: struct<@id: string>
child 0, @id: string
child 1, extract: struct<column: string>
child 0, column: string
child 2, transform: struct<jsonPath: string>
child 0, jsonPath: string
name: string
sameAs: string
to
{'@context': {'@language': Value('string'), '@vocab': Value('string'), 'citeAs': Value('string'), 'column': Value('string'), 'conformsTo': Value('string'), 'cr': Value('string'), 'rai': Value('string'), 'data': {'@id': Value('string'), '@type': Value('string')}, 'dataType': {'@id': Value('string'), '@type': Value('string')}, 'dct': Value('string'), 'equivalentProperty': Value('string'), 'examples': {'@id': Value('string'), '@type': Value('string')}, 'extract': Value('string'), 'field': Value('string'), 'fileProperty': Value('string'), 'fileObject': Value('string'), 'fileSet': Value('string'), 'format': Value('string'), 'includes': Value('string'), 'isLiveDataset': Value('string'), 'jsonPath': Value('string'), 'key': Value('string'), 'md5': Value('string'), 'parentField': Value('string'), 'path': Value('string'), 'recordSet': Value('string'), 'references': Value('string'), 'regex': Value('string'), 'repeated': Value('string'), 'replace': Value('string'), 'samplingRate': Value('string'), 'sc': Value('string'), 'separator': Value('string'), 'source': Value('string'), 'subField': Value('string'), 'transform': Value('string'), 'prov': Value('string')}, '@type': Value('string'), '@id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'conformsTo': Value('string'), 'url': Value('string'), 'sameAs': Value('string'), 'license': Value('string'), 'version': Value('string'), 'datePublished': Value('timestamp[s]'), 'dateModified': Value('timestamp[s]'), 'creator':
...
ue('string')}), 'citeAs': Value('string'), 'inLanguage': List(Value('string')), 'keywords': List(Value('string')), 'distribution': List({'@type': Value('string'), '@id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'contentUrl': Value('string'), 'encodingFormat': Value('string'), 'contentSize': Value('string'), 'sha256': Value('string')}), 'recordSet': List({'@type': Value('string'), '@id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'key': {'@id': Value('string')}, 'field': List({'@type': Value('string'), '@id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'dataType': Value('string'), 'source': {'fileObject': {'@id': Value('string')}, 'extract': {'column': Value('string')}, 'transform': {'jsonPath': Value('string')}}})}), 'rai:dataCollection': Value('string'), 'rai:dataCollectionType': List(Value('string')), 'rai:dataAnnotationProtocol': List(Value('string')), 'rai:dataBiases': List(Value('string')), 'rai:dataUseCases': List(Value('string')), 'rai:hasSyntheticData': Value('bool'), 'rai:personalSensitiveInformation': Value('string'), 'rai:dataLimitations': List(Value('string')), 'rai:dataSocialImpact': Value('string'), 'prov:wasDerivedFrom': {'@id': Value('string'), '@type': Value('string'), 'sc:name': Value('string'), 'sc:description': Value('string')}, 'prov:wasGeneratedBy': {'@id': Value('string'), '@type': Value('string'), 'sc:name': Value('string'), 'sc:description': Value('string')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
case_id: string
instruction: string
violation_type: string
violation_reason: string
correct_action: struct<action: string, coordinate: list<item: double>, description: string>
child 0, action: string
child 1, coordinate: list<item: double>
child 0, item: double
child 2, description: string
gt_action: struct<action: string, arguments: struct<text: string, coordinate: list<item: double>, time: double, (... 98 chars omitted)
child 0, action: string
child 1, arguments: struct<text: string, coordinate: list<item: double>, time: double, coordinate2: list<item: int64>, d (... 63 chars omitted)
child 0, text: string
child 1, coordinate: list<item: double>
child 0, item: double
child 2, time: double
child 3, coordinate2: list<item: int64>
child 0, item: int64
child 4, duration: int64
child 5, app: string
child 6, status: string
child 7, direction: string
img_path: string
layer: string
consequence_severity: string
risk_phase: string
auth_status: string
action_history: string
@id: string
rai:dataLimitations: list<item: string>
child 0, item: string
rai:hasSyntheticData: bool
url: string
rai:dataUseCases: list<item: string>
child 0, item: string
license: string
rai:dataAnnotationProtocol: list<item: string>
child 0, item: string
keywords: list<item: string>
child 0, item: string
citeAs: string
prov:wasDerivedFrom: struct<@id: string, @type: string, sc:name: string, sc:description: string>
child
...
0, @type: string
child 1, name: string
recordSet: list<item: struct<@type: string, @id: string, name: string, description: string, key: struct<@id: st (... 234 chars omitted)
child 0, item: struct<@type: string, @id: string, name: string, description: string, key: struct<@id: string>, fiel (... 222 chars omitted)
child 0, @type: string
child 1, @id: string
child 2, name: string
child 3, description: string
child 4, key: struct<@id: string>
child 0, @id: string
child 5, field: list<item: struct<@type: string, @id: string, name: string, description: string, dataType: string, s (... 118 chars omitted)
child 0, item: struct<@type: string, @id: string, name: string, description: string, dataType: string, source: stru (... 106 chars omitted)
child 0, @type: string
child 1, @id: string
child 2, name: string
child 3, description: string
child 4, dataType: string
child 5, source: struct<fileObject: struct<@id: string>, extract: struct<column: string>, transform: struct<jsonPath: (... 9 chars omitted)
child 0, fileObject: struct<@id: string>
child 0, @id: string
child 1, extract: struct<column: string>
child 0, column: string
child 2, transform: struct<jsonPath: string>
child 0, jsonPath: string
name: string
sameAs: string
to
{'@context': {'@language': Value('string'), '@vocab': Value('string'), 'citeAs': Value('string'), 'column': Value('string'), 'conformsTo': Value('string'), 'cr': Value('string'), 'rai': Value('string'), 'data': {'@id': Value('string'), '@type': Value('string')}, 'dataType': {'@id': Value('string'), '@type': Value('string')}, 'dct': Value('string'), 'equivalentProperty': Value('string'), 'examples': {'@id': Value('string'), '@type': Value('string')}, 'extract': Value('string'), 'field': Value('string'), 'fileProperty': Value('string'), 'fileObject': Value('string'), 'fileSet': Value('string'), 'format': Value('string'), 'includes': Value('string'), 'isLiveDataset': Value('string'), 'jsonPath': Value('string'), 'key': Value('string'), 'md5': Value('string'), 'parentField': Value('string'), 'path': Value('string'), 'recordSet': Value('string'), 'references': Value('string'), 'regex': Value('string'), 'repeated': Value('string'), 'replace': Value('string'), 'samplingRate': Value('string'), 'sc': Value('string'), 'separator': Value('string'), 'source': Value('string'), 'subField': Value('string'), 'transform': Value('string'), 'prov': Value('string')}, '@type': Value('string'), '@id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'conformsTo': Value('string'), 'url': Value('string'), 'sameAs': Value('string'), 'license': Value('string'), 'version': Value('string'), 'datePublished': Value('timestamp[s]'), 'dateModified': Value('timestamp[s]'), 'creator':
...
ue('string')}), 'citeAs': Value('string'), 'inLanguage': List(Value('string')), 'keywords': List(Value('string')), 'distribution': List({'@type': Value('string'), '@id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'contentUrl': Value('string'), 'encodingFormat': Value('string'), 'contentSize': Value('string'), 'sha256': Value('string')}), 'recordSet': List({'@type': Value('string'), '@id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'key': {'@id': Value('string')}, 'field': List({'@type': Value('string'), '@id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'dataType': Value('string'), 'source': {'fileObject': {'@id': Value('string')}, 'extract': {'column': Value('string')}, 'transform': {'jsonPath': Value('string')}}})}), 'rai:dataCollection': Value('string'), 'rai:dataCollectionType': List(Value('string')), 'rai:dataAnnotationProtocol': List(Value('string')), 'rai:dataBiases': List(Value('string')), 'rai:dataUseCases': List(Value('string')), 'rai:hasSyntheticData': Value('bool'), 'rai:personalSensitiveInformation': Value('string'), 'rai:dataLimitations': List(Value('string')), 'rai:dataSocialImpact': Value('string'), 'prov:wasDerivedFrom': {'@id': Value('string'), '@type': Value('string'), 'sc:name': Value('string'), 'sc:description': Value('string')}, 'prov:wasGeneratedBy': {'@id': Value('string'), '@type': Value('string'), 'sc:name': Value('string'), 'sc:description': Value('string')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
PhoneSafety
Safe, or Simply Incapable? Rethinking Safety Evaluation for Phone-Use Agents
Paper | Code and evaluation instructions | Croissant metadata
Dataset description
PhoneSafety evaluates the next action at a safety-critical phone state. Each case includes a Chinese task instruction, recent interaction history, a screenshot, and protocol-dependent safe and unsafe reference actions. The evaluation distinguishes Safe, Unsafe, and failure to realize either reference behavior (capability-failure rate, CFR).
The benchmark covers harmful-instruction refusal (Safety), user confirmation (Confirm), over-operation protection (OP), trap resistance (TR), and permission minimization (PM). Empirical coverage is Chinese Android interfaces.
Versions and public coverage
| Resource | Cases | Role |
|---|---|---|
| Research dataset | 700 | Dataset described in the paper |
| Clean-684 | 684 | Primary research analysis |
| Strict-652 analysis set | 652 | Reference-overlap sensitivity analysis |
| Public release | 640 | Downloadable cases in this repository |
| Public Clean analysis set | 627 | Public cases with automatically scorable strict-protocol references |
| Public Strict analysis set | 595 | Public cases after additionally excluding overlapping click targets |
Some research cases were removed from the public release to comply with company requirements. The public filenames retain 700, but each JSONL contains 640 records. Public-subset results therefore have different denominators from the full research tables.
The Strict analysis set is separate from the strict safety protocol. See the code repository's evaluation documentation for matching, exclusions, and coordinate conventions.
Files
| File | Contents |
|---|---|
phonesafety_700.jsonl |
640 cases with strict-protocol references; the primary evaluation setting |
phonesafety_700_minimal_protocol.jsonl |
The same 640 cases with supplementary minimal-protocol references |
screenshots.zip |
Screenshots corresponding to the public cases |
croissant.json |
Dataset metadata, responsible-AI information, provenance, and file checksums |
LICENSE |
CC BY-NC-SA 4.0 license text |
The strict and minimal files are two protocol views of the same public cases, not separate train/test splits. Each case_id identifies the screenshot basename in the archive. Full action parameters are retained in the JSONL.
Record fields
| Field | Meaning |
|---|---|
case_id |
Unique public case identifier |
instruction |
Chinese user instruction |
action_history |
Recent interaction history |
img_path |
Original screenshot path; use its basename in the screenshot archive |
violation_type |
Scenario family: Safety / Confirm / OP / TR / PM |
violation_reason |
Explanation of the annotated safety boundary |
correct_action |
Protocol-dependent safe reference action, with available target or text parameters |
gt_action |
Protocol-dependent unsafe reference action, with available target or text parameters |
layer |
Task-level or step-level risk |
The strict file additionally includes diagnostic fields such as consequence_severity, risk_phase, and auth_status. The minimal file includes flipped and the original strict-protocol references. These supplementary fields are not present uniformly across both files.
Download and evaluation
The public files can be downloaded without a personal access request. To prepare the data and screenshots in the layout expected by the scorer:
git clone https://github.com/tangzhy/PhoneSafety.git
cd PhoneSafety
python3 setup_data.py
For saved predictions, run:
python3 inference/evaluate.py \
--predictions outputs/your_model_strict.jsonl \
--benchmark data/phonesafety_700.jsonl \
--analysis-set clean \
--output outputs/your_model_strict_scores.json
The offline scorer uses Python's standard library. Model inference requires the additional dependencies and endpoint configuration described in the code README.
The scorer applies 22 versioned coordinate export corrections to the current public strict JSONL in memory after checking the dataset checksum and screenshot dimensions. Downloaded files are not changed. Use the published scorer to reproduce its matching conventions rather than interpreting every stored coordinate as having the same unit.
Authorship and provenance
The dataset was created by the PhoneSafety paper authors, affiliated with Tencent Hunyuan, The Chinese University of Hong Kong, Shenzhen, and Tsinghua University. Full authorship is provided in the citation below.
The study collected 4,512 human-executed Android trajectories with approximately 75,000 steps across more than 130 Chinese apps. Safety-critical states were screened, reviewed, and expanded to form the research dataset. Human case-validity review occurred before reference construction. Safe and unsafe references were inherited from earlier review or generated by Method B using the task, history, screenshot context, and active protocol. Machine-generated reference annotations are the reason rai:hasSyntheticData is true in the metadata.
Historical verification was sequential quality control, not independent dual annotation. The paper separately reports a follow-up blinded human agreement study on selected agent outputs (Appendix M).
Intended use and limitations
- Intended use is offline research and diagnostic safety evaluation at phone decision points under an explicit protocol.
- The benchmark is not a deployment safety certificate and does not evaluate complete online trajectories, recovery, or cumulative long-horizon risk.
- Chinese Android interfaces, app selection, and authorization conventions shape the case distribution. English-interface, other-OS, unseen-app, updated-app, and temporal transfer are not established.
- Paired references do not enumerate every acceptable action. CFR is an operational matching category that can reflect language and app familiarity, perception, grounding, or output formatting; it does not identify a single latent inability.
- The public subset omits 60 research cases, limiting exact reproduction of full research-set tables from public data alone.
- Screenshots and histories include safety-sensitive contexts such as contact, payment, and permission interfaces. The paper describes release safeguards in its checklist. The benchmark should be studied as offline evaluation data rather than used to act on depicted accounts, identities, or transactions.
- Safety-critical cases can help diagnose agent weaknesses but can also be misused to probe those weaknesses. Scores should be interpreted together with the protocol, reference limitations, and application scope.
License
This release is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). The license requires attribution, limits use to noncommercial purposes, and includes share-alike conditions. See the complete license text for its terms.
Contact
Maintainer: Zhengyang Tang, zhengyang.tzy@gmail.com.
Citation
@article{tang2026phonesafety,
title={Safe, or Simply Incapable? Rethinking Safety Evaluation for Phone-Use Agents},
author={Zhengyang Tang and Yi Zhang and Chenxin Li and Xin Lai and Pengyuan Lyu and Yiduo Guo and Weinong Wang and Junyi Li and Yang Ding and Huawen Shen and Zhengyao Fang and Xingran Zhou and Liang Wu and Fei Tang and Sunqi Fan and Shangpin Peng and Zheng Ruan and Anran Zhang and Benyou Wang and Chengquan Zhang and Han Hu},
journal={arXiv preprint arXiv:2605.07630},
year={2026}
}
- Downloads last month
- 56