Benchmark
AgentDojo
Test whether tool-using agents complete tasks despite prompt-injection attacks.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The original benchmark includes 97 realistic user tasks and 629 security cases involving workplace, banking, and travel tools.
Methodology
Measure task utility, utility under attack, and attack success for a specific model, attack, and defense combination.