Wallbreaker Turns AI Jailbreaking Into a Fully Automated Attack Agent
An open-source red-team harness turns jailbreaking into a repeatable pipeline with 222 transforms, HarmBench batteries, and autonomous attack loops.
Takeaways
- Wallbreaker is an open-source terminal agent for automated LLM red-teaming, released under AGPL-3.0 on GitHub.
- Ships with 222 obfuscation transforms via the vendored P4RS3LT0NGV3 engine and a Python fallback.
- Uses HarmBench's 400 behaviors across 7 categories instead of cherry-picked prompts.
- Separates attacker brain, target model, and judge model, all configurable via TOML.
- Includes PAIR, Crescendo, best-of-N, many-shot, prefill, and multimodal image-chain attacks as callable tools.
- Headless CLI mode exports findings JSON and can fail CI builds when a bypass is found.
Red-teaming large language models has mostly been a cottage industry of clever prompts pasted into chat windows. Wallbreaker tries to industrialize it as a Claude-Code-style terminal agent that reasons in a loop, calls tools, mutates payloads, and grades its own results against a benchmark battery until something breaks or a human is asked to intervene.
You talk to Wallbreaker the way you would talk to Claude Code, and it reasons and calls tools in a loop. The backend is provider-agnostic, running on OpenRouter, the Z.AI GLM coding plan, the local Claude Code CLI, a local server, or any OpenAI- or Anthropic-compatible API. It ships with the Parseltongue transform engine, the L1B3RT4S jailbreak library, the HarmBench behavior benchmark, automated attack loops (PAIR/TAP, Crescendo, best-of-N), a from-scratch persona author, native-format target mimicry drawn from a leaked system-prompt corpus, a multimodal image-edit attack channel, an LLM judge, and reliability validation.
Three roles, one config
The tool separates three roles in a config file: the attacker brain, the target model under attack, and a judge model that grades replies. That separation is the point. You can pit GLM-4.6 against Claude, use GPT-4o-mini as an impartial grader, and swap any of them without touching code.