Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
52 lines (41 loc) · 2.48 KB

CONTRIBUTING.md

File metadata and controls

52 lines (41 loc) · 2.48 KB
Raw

Contributing to Wallbreaker

Thanks for helping make LLM red-teaming better. By contributing you agree your work is licensed under the project's AGPL-3.0-or-later license, and that you'll use the tool within the bounds described in SECURITY.md.

Dev setup

python -m venv .venv
. .venv/bin/activate
pip install -e ".[dev]"        # add [barcodes] / [stego] for those tools
wallbreaker check              # validate config (profiles, keys, target, judge)
pytest -q                      # full suite must stay green

Architecture (where things live)

  • wallbreaker/providers/ normalize OpenAI + Anthropic wire formats to one event stream (agent/messages.py); agent/loop.py is protocol-agnostic.
  • wallbreaker/tools/ — each tool is a module exposing register(registry); add the module name to tools/__init__.py to wire it in.
  • wallbreaker/transforms/ — pure encode/decode functions, indexed in transforms/__init__.py with a lossy flag.
  • wallbreaker/presets.py — curated single-shot jailbreak templates.
  • wallbreaker/tui/ — the Textual terminal UI (theme in theme.py, chrome in header.py/sidebar.py/widgets.py, layout in app.tcss).
  • dashboard/ — FastAPI backend + React/Svelte web dashboard.

House rules

  • No comments or emoji in code. Use short docstrings where they add real value.
  • Presets are .format()-filled — keep literal {/} out of templates (use pipes or brackets for dividers); {request} must be the only brace token. A test enforces this.
  • New tools register into a LOCAL ToolRegistry in their own test, not build_registry().
  • Transforms: mark lossy ones lossy=True; lossless ones must round-trip exactly.
  • Add a new tool by writing register(registry) and appending the module name to the tuple in tools/__init__.py (a serial collision hub — one edit at a time).
  • Run pytest -q before opening a PR; the suite is the contract.

Adding a jailbreak technique

Most techniques land as either a preset (presets.py), a transform (transforms/), or a tool (tools/), each following the conventions of its category. Label generic academic techniques honestly (cite the paper) rather than overclaiming novelty.

Scope reminder

Contributions that make the tool better at finding and reporting weaknesses are welcome. Contributions whose only purpose is to maximize real-world harm (e.g. shipping weaponized payloads with no evaluation value) are not.