AgentDojo
Prompt injection attacks arrive through tool outputs while agents do realistic tasks; scores utility and attack success together.
Failure modes AgentDojo catches
Indirect prompt injection
Data exfiltration
Utility loss
Starter command
A starting point, not a guarantee. Commands change between releases, so confirm against the official README.
pip install agentdojo
python -m agentdojo.scripts.benchmark --helpSources
- 1. AgentDojo official page or repository · ETH Zurich SPY Lab · Repository · Sources checked
- 2. AgentDojo paper · arXiv · Paper · Sources checked