datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
moltbook-agent-social-ai-prompt-injection-dataset
Moltbook Agent-Social AI Prompt Injection Dataset
207,391 items — 77,469 posts and 129,922 comments — from Moltbook, a social network whose users are AI agents.
Scanned for indirect prompt-injection patterns using the taxonomy of Greshake et al. (2023). The full raw corpus is included, so you can ignore my analysis entirely and do your own.
These are keyword-matched candidates, not verified attacks. An agent discussing prompt injection matches the same words as one performing… See the full description on the dataset page: https://huggingface.co/datasets/DavidTKeane/moltbook-agent-social-ai-prompt-injection-dataset.geo-injection-rag-attack-data
Can It Reach the Generator? Investigating the Survival of GEO Prompt-Injection Attacks in Realistic RAG Settings
This dataset contains the prompt-injection attack data presented in the paper Can It Reach the Generator? Investigating the Survival of Prompt-Injection Attacks in Realistic RAG Settings.
The dataset is used to… See the full description on the dataset page: https://huggingface.co/datasets/Euanyu/geo-injection-rag-attack-data.prompt_injectionsdataset-injection-scan-study
Dataset Injection Scan — open study of popular HF datasets
from datasets import load_dataset
ds = load_dataset("fevziegeyurtsevenler/dataset-injection-scan-study")
Results of scanning 17,000 rows across 6 popular public instruction/prompt datasets for
smuggled prompt-injection with hf-dataset-scan
(invisible Unicode, injection phrasing EN+TR, exfil URLs).
Headline: no smuggled injection found
Dataset
Rows
Flagged
High
Med
Low
tatsu-lab/alpaca
3,000
0… See the full description on the dataset page: https://huggingface.co/datasets/fevziegeyurtsevenler/dataset-injection-scan-study.multiturn-injection-detection
Multi-Turn Distributed Prompt Injection Detection Dataset
Dataset Description
27,180 synthetic multi-turn conversations (18,754 train / 3,296 val / 5,130 test) designed for training and evaluating temporal prompt injection detectors. Each conversation consists of 6-9 user turns with assistant responses.
Shared-Prefix Design
Every attack conversation is paired with a benign conversation that shares identical opening turns. A conversational prefix of 3-5 user… See the full description on the dataset page: https://huggingface.co/datasets/rockCO78/multiturn-injection-detection.