jailbreak-corpus-mini-py

Small local jailbreak and prompt-injection fixture set for tests. Ships a JSON corpus of canonical attack prompts (instruction override, prompt leak, excessive agency, encoding tricks, role hijack) with category + severity labels you can pivot on. Zero runtime dependencies.

Python port of @mukundakatta/jailbreak-corpus-mini.

Install

pip install jailbreak-corpus-mini-py

Usage

from jailbreak_corpus_mini import (
    load_corpus,
    get_jailbreak_fixtures,
    fixture_texts,
    Fixture,
)

# Full corpus (each entry is a Fixture: {id, prompt, category, severity, text}).
corpus = load_corpus()
len(corpus)                    # > 5
corpus[0].prompt               # canonical attack string
corpus[0].category             # e.g. "prompt_injection"
corpus[0].severity             # "low" | "medium" | "high"

# Filter to a single category (matches JS sibling's getJailbreakFixtures).
get_jailbreak_fixtures(category="prompt_injection")

# Just the raw strings (handy for hooking into a guardrail eval loop).
texts = fixture_texts(severity="high")

Use it in a guardrail test

import pytest
from jailbreak_corpus_mini import load_corpus

def my_guardrail(prompt: str) -> bool:
    """Return True if prompt should be blocked."""
    ...

@pytest.mark.parametrize("fixture", load_corpus())
def test_guardrail_blocks_known_attacks(fixture):
    assert my_guardrail(fixture.prompt), (
        f"missed {fixture.category}/{fixture.severity}: {fixture.prompt!r}"
    )

API

The fixture corpus ships as corpus.json inside the package and is loaded once on import. Patch the file (or call load_corpus_from(path)) to use your own.

See the JS sibling's README for the full design notes.