Datasets:
license: mit
language:
- en
task_categories:
- text-classification
tags:
- ai-safety
- llm-security
- prompt-injection
- agentic-social-networks
- owasp-llm-top-10
- Moltbook
size_categories:
- 1K<n<10K
pretty_name: MoltSafe-10K
MoltSafe-10K
MoltSafe-10K contains safety annotations for 10,000 posts and comments from Moltbook, an agentic social network
where autonomous agents communicate with one another. The underlying corpus is
AIcell/moltbook-data on the
Hugging Face Hub. Every node has a binary safety verdict, a severity level, a
malicious-intent taxonomy, and OWASP GenAI risk codes.
We release the annotations used in the development our study to promote research into Moltbook security. The dataset contains the labels only, without the associated malicious text. In order to obtain the associated text, the reader will need to download the original dataset, and then, based on the post and comment ids to get the malicious text. The text of a certain post does not include the associated comments.
Schema
| Column | Description |
|---|---|
node_id |
post:<uuid> or comment:<uuid> |
node_type |
post or comment |
verdict |
safe, unsafe |
severity |
0 to 5 |
risk_taxonomies |
up to 3 per node, unsafe nodes only |
owasp_risk_codes |
up to 3 per node |
Severity scale
Levels are anchored on operational impact in an autonomous-agent environment.
| Severity | Label | Description |
|---|---|---|
| 0 | Safe | No harmful intent or effect |
| 1 | Weak signal | Visible but weak unsafe signal, limited impact |
| 2 | Low operational impact | Harmful, abusive or deceptive, low direct operational impact |
| 3 | Plausible operational risk | Manipulative or coordinated harmful influence carrying plausible operational risk, but no clear compromise |
| 4 | Localized compromise | Unauthorized execution, sensitive-data exposure, or single-agent hijack |
| 5 | Systemic compromise | Self-propagating or cascading multi-agent compromise, or unbounded resource abuse |
Risk taxonomy
Categories of malicious intent, assigned to unsafe nodes only. They are not mutually exclusive.
prompt_injection, jailbreak_or_safety_bypass,
sensitive_information_extraction_or_leakage, engagement_manipulation,
behavior_or_memory_poisoning, social_engineering_attacks,
harmful_or_abusive_content, financial_scams_or_fraud,
misinformation_or_hallucination_cascade, unsafe_execution_directive, other
OWASP GenAI Top-10 codes LLM01 through LLM10 are assigned as well.
Construction
Starting from the original dataset, the following steps were employed to generate the dataset:
- Hash-deduplicated the original corpus.
- Language tagged the nodes using the Lingua library in high-accuracy mode, keeping only English-dominant content. A node was kept when more than 50% of its characters fell in English-tagged windows.
- Sampled 10,000 nodes without replacement, after removing blank records.
- Annotated them with GPT-5.5 at reasoning effort high.
For more details, please visit [link to paper]
The code used to develop this dataset and reproduce the experiments is publicly available on GitHub.
Data Provenance and Licensing
MoltSafe-10K releases only the annotations and source identifiers used in this study and does not redistribute the associated Moltbook text. The original posts and comments can be retrieved separately from AIcell/moltbook-data using the provided identifiers.
The MoltSafe-10K license applies to the released annotations and accompanying materials only and does not extend to the original source text. Users are responsible for complying with the applicable terms and data-protection requirements of the source dataset.
Citation
Paper under review. Citation to be added.