AgentDojo

Evals

Benchmark / dataset

Dynamic environment for prompt-injection attacks and defenses on tool-using agents.

Category
Prompt injection & agent security
Insurance exposure
Prompt injection
Publisher
ETH Zurich (SPY Lab)
Publisher type
Academic
License
MIT
Status
Maintained (<12mo)
Last commit
2026-06-02
Stars
925
Notes
Used by UK/US AISIs.

View source on GitHub →

Wanich Risk Lab, 2026.