MLCommons Jailbreak Benchmark v1.0

The MLCommons Jailbreak Benchmark v1.0 is a benchmark for evaluating how well large language models (LLMs) withstand jailbreak attacks. Modern AI systems are designed to refuse hazardous requests. A jailbreak is a prompt crafted to bypass these safeguards and obtain outputs the system would normally refuse to provide. The benchmark gives an end-to-end method for measuring how resilient AI systems are to single-turn, text-based jailbreaks. It was published in October 2026 by MLCommons, an industry and academic consortium, with authors led by the University of Warwick.

The benchmark applies representative jailbreak techniques, drawn from published research, practitioner reports and community submissions. These are applied to prompts covering eleven hazard categories, including violent crime, indiscriminate weapons, hate and child sexual exploitation. It evaluates known classes of jailbreak techniques rather than developing new attacks.

Each AI system is first tested with the AILuminate Safety Benchmark v1.4, which provides a safety baseline. The same system is then tested with jailbreak versions of the prompts, and the two results are compared to show how much the attacks weaken its safety. The benchmark measures how often each system produces content that violates the AILuminate Assessment Standard. A “resilient” system seldom does, even when faced with emerging jailbreaks. The pipeline combines criteria for selecting systems and attacks, human annotation, calibration of automated evaluators, scoring, grading and risk-calibrated disclosure of results.

Auto-discovered on 2026-10-07 by OECD Catalogue Automation