evaluation-environment · Documentary review

AgentDojo

Dynamic environment for evaluating attacks and defences for tool-using agents.

Open interactive record →
OriginETH Zurich SPY Lab and Invariant LabsTopicsAutonomy & agents · SafeguardsStatusReviewed

Can support

Utility-security tradeoffs against AgentDojo's prompt-injection scenarios under the exact agent and defense configuration.

Cannot support by itself

All prompt-injection vectors, live tool ecosystems, supply-chain attacks, or real-world incident frequency.

Decision use

Best used for

Comparing agents and defenses on reproducible indirect prompt-injection scenarios.

Not enough for

Claims that a production agent is secure against prompt injection.

Original sources