AgentHarm & AgentDojo

What it measures

Evaluates safety and prompt-injection defenses specifically in multi-step tool-calling agents (fraud, cybercrime, data exfiltration).

Reading the measurement

Separate satisfying instructions from factual accuracy, human preference and safe behavior. Inspect strict all-constraint success as well as average criterion credit. Judge choice and adversarial prompt selection can materially change the result.

Published evaluation

AgentDojo targeted attack success; AgentDojo · agent comparison. The source’s task split, model configuration and scoring units are retained.

Where this fits in an evaluation suite

Format compliance, preference, truthfulness, and safety require separate measurements. Keep desirable writing style from masking factual errors.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

AgentDojo targeted attack success; AgentDojo · agent comparison. The source’s task split, model configuration and scoring units are retained.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.