AgentDojo is the de-facto standard benchmark for prompt-injection and agent security: a dynamic environment of realistic tool-using tasks paired with adversarial injection attacks, measuring both utility and robustness. From Tramèr's SPY Lab; NeurIPS 2024 Datasets & Benchmarks and a SafeBench first prize. Adopted into UK AISI's inspect_evals, used in US/UK AISI joint testing, and the substrate for Google's CaMeL defense. Co-authors later founded Invariant Labs (acquired by Snyk, 2025).

Evaluation Details

Domains 3
Domains: prompt injection, agent security, tool use
evalbenchmarksecurityagenticsafety