Multi-Turn Human Jailbreaks (MHJ)

Evals

Benchmark / dataset

Human multi-turn jailbreaks that beat defenses robust to single-turn attacks.

Category
Jailbreak & misuse
Insurance exposure
Regulatory/bias
Publisher
Scale AI
Publisher type
Company (open source)
License
unknown
Status
Not on GitHub
Last commit
unknown
Notes
Dataset on HF; not on GitHub.

View source on GitHub →

Wanich Risk Lab, 2026.