license: mit
language:
- en
tags: - prompt-injection
- jailbreak
- llm-security
- red-team
- safety
- testing
size_categories: - n<1K
configs: - config_name: default
data_files:- split: train
path: data.jsonl
- split: train
prompt-injection-patterns-extended
30 representative prompt-injection patterns across 10 categories. Built for testing input filters, agent defenses, and red-team eval pipelines. Pairs nicely with the smaller jailbreak-corpus-mini fixture set.
Categories
| Category | Count | Example |
|---|---|---|
instruction-override |
2 | "Ignore all previous instructions..." |
role-confusion |
2 | "From now on you are DAN..." |
obfuscation |
4 | base64 / ROT13 / homoglyph attacks |
indirect |
4 | Hidden in RAG pages or HTML comments |
tool-abuse |
2 | Tries to call dangerous tools with bad args |
data-exfiltration |
2 | Asks to leak prior chat / env vars |
jailbreak |
2 | Classic hypothetical / authority framing |
context-pollution |
2 | Fakes past agreements |
auth-spoofing |
2 | Fake admin tokens |
multi-step |
2 | Chained nested instructions |
delimiter-injection |
2 | Fakes system/user XML or markdown delimiters |
language-switch |
2 | Same attack in non-English |
social-engineering |
2 | Sob-story manipulation |
Schema
{
"id": "string",
"category": "string",
"text": "string", // the attack payload
"severity": "critical | high | medium" // analyst-rated risk
}
Quickstart
from datasets import load_dataset
ds = load_dataset("mukunda1729/prompt-injection-patterns-extended", split="train")
critical = [r for r in ds if r["severity"] == "critical"]
print(f"{len(critical)} critical patterns")
Notes
These are defensive fixtures — patterns to test your filters and agent guardrails against. Do not use them to probe production systems you don't own.
Related
License
MIT.