Wallbreaker is an AI red-teaming harness for authorized LLM security testing. It provides a configurable, loop-driven agent that automates attack strategies using advanced techniques. The tool benefits security researchers by offering a comprehensive suite for evaluating LLM robustness and identifying vulnerabilities in a standardized, reproducible manner.

How It Works

Wallbreaker operates as a terminal agent executing a reasoning and tool-calling loop against LLMs. Its backend supports diverse API providers (OpenAI, Anthropic, OpenRouter) and local environments. Key components include the Parseltongue transform engine for obfuscation, L1B3RT4S jailbreak library, and HarmBench for standardized testing. Autonomous attack loops (PAIR/TAP, Crescendo) enable iterative refinement until success or operator intervention.

Quick Start & Requirements

• Installation: Use Python virtual environments (python -m venv .venv, . .venv/bin/activate), then pip install -e ".[dev]".
• Prerequisites: Python 3.x, Node.js (for full Parseltongue). A config.toml file is required for API keys and endpoint configurations.
• Setup: Copy config.example.toml to config.toml and populate credentials. Run wallbreaker check for validation.
• Dashboard: Additional setup: pip install -e ".[dashboard]", cd wallbreaker/dashboard/web && npm install && npm run build && cd -.
• Links: CONTRIBUTING.md and SECURITY.md provide detailed setup and policies.

Highlighted Details

• Parseltongue Engine: Offers native Python (50+ transforms) and full Node.js (222 transforms) for advanced string manipulation and obfuscation.
• Autonomous Attack Loops: Implements iterative strategies like PAIR/TAP and Crescendo to maximize bypass success rates.
• HarmBench Integration: Utilizes the HarmBench benchmark (400+ behaviors) for standardized LLM safety evaluations.
• Persona Authoring & Native Mimicry: Tools to author system prompts via ENI method and mimic target LLM formats using a leaked system prompt corpus.
• Multimodal Attacks: Supports image-based attacks, including editing and decomposing refused generations.

Maintenance & Community

Setup, architecture, and contribution guidelines are detailed in CONTRIBUTING.md. Security policy is in SECURITY.md. No explicit community channels (e.g., Discord, Slack) are mentioned.

Licensing & Compatibility

Licensed under AGPL-3.0-or-later. This strong copyleft license requires that modified versions, including network/hosted services, must make their complete corresponding source code available under the same license. Third-party corpora are fetched at runtime.

Limitations & Caveats

Strictly for "authorized testing only." Full Parseltongue capabilities require Node.js. The AGPL-3.0 license imposes significant obligations for distributed or hosted use.