dmilush/shieldlm-prompt-injection

ShieldLM Prompt Injection Dataset

A unified prompt injection detection dataset with 54,162 samples spanning three attack categories: direct injection, indirect injection, and jailbreak. Curated from 11 source datasets with a 3-level hierarchical label schema.

Dataset Description

Purpose

Training and evaluating prompt injection classifiers for production deployment. Designed to address gaps in existing datasets:

• Indirect injection coverage (via InjecAgent tool-embedded attacks)
• Jailbreak techniques (via TrustAIRLab in-the-wild prompts + jackhhao classification)
• Application-structured benign data (prevents format-based false positives)
• Sensitive-topic stress tests (via JailbreakBench — topics != techniques)

Statistics

Metric	Value
Total samples	54,162
Train / Val / Test	37,913 / 8,124 / 8,125
Benign	35,197 (65.0%)
Attack	18,965 (35.0%)
Languages	8 (en, fr, es, it, de, pt, ro, ca)
Sources	11 datasets

Category breakdown:

Category	Count	%
benign	35,197	65.0%
direct_injection	16,893	31.2%
indirect_injection	1,054	1.9%
jailbreak	1,018	1.9%

Label Schema

Level	Field	Values
1 (Binary)	label_binary	0 (BENIGN), 1 (ATTACK)
2 (Category)	label_category	benign, direct_injection, indirect_injection, jailbreak
3 (Intent)	label_intent	goal_hijacking, data_exfiltration, financial_harm, ...

Source Datasets

Source	License	Category	Samples
alespalla/chatbot_instruction_prompts	Apache-2.0	benign (conversational)	24,804
reshabhs/SPML_Chatbot_Prompt_Injection	CC-BY-4.0	direct_injection + benign	15,913
xTRam1/safe-guard-prompt-injection	Apache-2.0	direct_injection + benign	8,118
TrustAIRLab/in-the-wild-jailbreak-prompts	CC-BY-NC-SA-4.0	jailbreak	1,002
Harelix/Prompt-Injection-Mixed-Techniques-2024	Apache-2.0	direct_injection + benign	987
yanismiraoui/prompt_injections	Apache-2.0	direct_injection (multilingual)	974
deepset/prompt-injections	Apache-2.0	direct_injection + benign	546
InjecAgent (UIUC)	MIT	indirect_injection + benign	1,054
jackhhao/jailbreak-classification	MIT	jailbreak + benign	531
JailbreakBench/JBB-Behaviors	MIT	benign (FP stress test)	200

Note: Harelix data recovered from ahsanayub/malicious-prompts (original dataset removed from HuggingFace).

Key Design Decisions

1. JailbreakBench Goals are labeled BENIGN — they describe harmful topics, not injection techniques. Used as a false-positive stress test.
2. No isolated attacker payloads from InjecAgent — context determines injection (PromptShield insight).
3. Benign data includes application-structured samples — clean tool responses from InjecAgent prevent the classifier from learning "JSON format = attack."
4. In-the-wild jailbreaks — real DAN, persona switching, and role-play attacks collected from Reddit, Discord, and jailbreak forums.

Fields

Field	Type	Description
id	string	Deterministic unique ID: {source}_{index}_{hash}
text	string	Text to classify
label_binary	int	0 (BENIGN) or 1 (ATTACK)
label_category	string	One of 4 categories
label_intent	string (nullable)	Fine-grained attack intent
source	string	Origin dataset
language	string	ISO 639-1 language code
context	string (nullable)	System prompt or user instruction
metadata	dict	Source-specific metadata

Splits

Split	Samples	Purpose
train	37,913	Model training (70%)
validation	8,124	Threshold calibration and model selection (15%)
test	8,125	Final evaluation (15%)

Stratified by label_category, random seed 42.

Intended Use

• Training prompt injection detection classifiers
• Benchmarking detection systems at low-FPR operating points
• Research on adversarial robustness of LLM safety filters

Limitations

• English-dominant: >98% English; multilingual samples limited to 7 other languages
• Text-only: No multimodal or visual prompt injection
• Synthetic benign tool responses: Generated by stripping injections from InjecAgent
• Static benchmark: Does not capture evolving attack techniques
• No multi-turn: All samples are single-turn

Citation

@software{shieldlm2026,
  author = {Milushev, Dimiter},
  title = {ShieldLM: Unified Prompt Injection Detection Dataset},
  year = {2026},
  url = {https://github.com/dvm81/shieldlm}
}

License

MIT (this curation). Source datasets retain their original licenses (see table above). Note: TrustAIRLab/in-the-wild-jailbreak-prompts uses CC-BY-NC-SA-4.0; all other sources are Apache-2.0, CC-BY-4.0, or MIT.