AI & LLM Security — Türkçe-first
Collection
Open datasets, tools and demos for AI/LLM & agent security — multilingual, Turkish-first. By AltaySec. • 25 items • Updated
Datasets:
Tasks:
Text Classification
Formats:
json
Size:
< 1K
Tags:
prompt-injection
training-data-poisoning
supply-chain-security
invisible-unicode
llm-security
ai-security
License:
dataset stringclasses 6
values | rows_scanned int64 2k 3k | rows_flagged int64 0 10 | flag_rate float64 0 0 | high int64 0 0 | medium int64 0 0 | low int64 0 11 | by_family stringclasses 5
values |
|---|---|---|---|---|---|---|---|
tatsu-lab/alpaca | 3,000 | 0 | 0 | 0 | 0 | 0 | {} |
yahma/alpaca-cleaned | 3,000 | 0 | 0 | 0 | 0 | 0 | {} |
databricks/databricks-dolly-15k | 3,000 | 4 | 0.0013 | 0 | 0 | 4 | {"hidden-unicode": 4} |
fka/awesome-chatgpt-prompts | 2,000 | 9 | 0.0045 | 0 | 0 | 9 | {"injection-phrase": 4, "hidden-unicode": 5} |
HuggingFaceH4/no_robots | 3,000 | 3 | 0.001 | 0 | 0 | 3 | {"hidden-unicode": 3} |
Open-Orca/OpenOrca | 3,000 | 10 | 0.0033 | 0 | 0 | 11 | {"hidden-unicode": 11} |
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 | 0 | 0 | 0 |
| yahma/alpaca-cleaned | 3,000 | 0 | 0 | 0 | 0 |
| databricks/databricks-dolly-15k | 3,000 | 4 | 0 | 0 | 4 |
| fka/awesome-chatgpt-prompts | 2,000 | 9 | 0 | 0 | 9 |
| HuggingFaceH4/no_robots | 3,000 | 3 | 0 | 0 | 3 |
| Open-Orca/OpenOrca | 3,000 | 10 | 0 | 0 | 10 |
| Total | 17,000 | 26 | 0 | 0 | 26 |
0 decoded Tags-block instructions · 0 bidi overrides · 0 genuine exfil URLs. All 26 low flags were benign on inspection (a prompts dataset naturally contains prompt-like phrasing; the rest were zero-width/BOM formatting). The invisible-injection threat to training data is real in theory but essentially absent in these corpora today — an honest, reproducible counter-hype baseline.
Full method + interpretation: FINDINGS.md. Reproduce or scan your own data with the tool: 👉 github.com/fevziegeyurtsevenler/hf-dataset-scan
Related
🕵️ uncloak · 🌿 skills-in-the-wild · 🏁 guardrail-arena · 🌐 AltaySec
Citation
@misc{yurtsevenler2026hfdatasetscan,
title={hf-dataset-scan: Scanning Datasets for Smuggled Prompt-Injection},
author={Yurtsevenler, Fevzi Ege}, year={2026}, publisher={AltaySec},
howpublished={\url{https://github.com/fevziegeyurtsevenler/hf-dataset-scan}}
}
Apache-2.0 · by AltaySec.
- Downloads last month
- 51