Jailbreaker is a local offline evaluation harness for testing chatbot and agent systems against jailbreak, prompt-injection, and related safety failures.

Launch write-up: Jailbreaker: LLM Jailbreak Testing You Can Actually Repeat

cp .env.example .env
docker compose up -d --build frontend backend database

Open http://localhost:3000.

Read the full setup guide in docs/getting-started.md and the security model in docs/security-model.md.

• Getting started
• Configuration
• Architecture
• Security model
• Local development
• Release process

Target, Attacker, and Judge system prompts are independent run configuration and are not stored in reusable LLM vault profiles. Target prompts are used directly; Attacker and Judge prompts are prepended to Jailbreaker's required technique and scoring instructions.

Judge evaluation criteria can use the existing standard behavior, explicit success and failure outcomes, or a complete custom rubric. Custom criteria keep the platform's 1-10 score contract, with 7 as the minimum successful score, and are shared by Dashboard runs, Experiment Design, and experiment retries.

Experiment datasets may optionally include a gold_answer on individual prompt objects. When present, Jailbreaker supplies it to the Judge as a trusted factual reference; prompts without one retain the standard Judge behavior. Gold answers are never added to Target or Attacker messages. Because gold_answer is treated as trusted Judge configuration, import experiment datasets only from sources you trust and do not place instructions or secrets in that field.

See CONTRIBUTING.md for issue, pull request, and development guidance.

Contributors are acknowledged in release notes and pull requests.

This project is currently maintained by Neeraj Gupta.