WallBreaker: Open‑Source AI Red‑Team Harness for One‑Click Automated LLM Jailbreak

WallBreaker is an open‑source AI red‑team harness maintained by SingulCore (GitHub: JailbrokenAI/wallbreaker) that automates security testing of large language models (LLMs) using an agent‑driven approach.

Core features include an autonomous attack loop that repeatedly modifies prompts until the model “surrenders” (via finish() or ask_operator()), the Parseltongue transformation engine with 59 built‑in text‑morphing techniques (expandable to 222 with the P4RS3LT0NGV3 engine), multiple advanced attack modules such as CiperChat (ICLR 2024), Skeleton_Key (Russinovich 2024), Persuasion Attack (Zeng 2024), DRAttack (Li 2024) and ICA (Wei 2023), multi‑model support (OpenRouter, Z.AI GLM Coding Plan, OpenAI API, Anthropic API, local Claude Code CLI), and an integrated LLM Judge that scores attacks against 400 HarmBench behaviors.

Performance on DeepSeek – according to the project documentation, WallBreaker achieves an average jailbreak success rate of about 60 % on DeepSeek‑Chat and 93 % on Claude Opus 5 (full‑domain cracking), demonstrating that even models considered highly protected can be compromised by automated pipelines.

Installation & usage – the tool can be installed with a simple Git clone, virtual‑environment setup, and pip install command. Configuration is done via config.toml, where API endpoints and keys for the target model and the judge are specified. After verification with wallbreaker check, attacks are launched with wallbreaker.

git clone https://github.com/JailbrokenAI/wallbreaker
cd wallbreaker
python -m venv .venv
. .venv/bin/activate
pip install -e "[dev]"
cp config.example.toml config.toml  # fill in your API keys

Red‑team perspective – the automation turns manual jailbreak attempts that previously took days into an “assembly‑line” that can generate bypasses within hours, enabling continuous security assessment of LLMs.

Blue‑team perspective – the 60 % success rate highlights existing blind spots in model defenses; teams can use the open‑source tool to probe their own systems, identify vulnerabilities, and harden defenses.

Broader AI‑security ecosystem – tools like WallBreaker are essential for driving defensive improvements, as pressure testing reveals weaknesses that would otherwise remain hidden.