← Evals

Evaluation profile

AgentDojo

2sub-evals
0.953%total index weight
1components

Within-component eval weight: Responsible agency 6.36%.

Model score (see each sub-eval direction)Predicted score

About this eval

Robust and useful delegated-agent behavior under prompt injection.

Included in the behavior ranking.

Sub-evals

MeasureComponentDirectionTotal index weightWithin-component weight
targeted_attack_success_rateagentdojo/agentdojo.csv:targeted_attack_success_rateMeasures whether prompt injections make a tool-using agent carry out the attacker’s target action.responsible_agency_control:1.000agentdojo/agentdojo.csvLower is better0.715%Responsible agency 4.77%
utility_under_attackagentdojo/agentdojo.csv:utility_under_attackMeasures whether a tool-using agent still completes the user’s intended task while an injection attack is present.responsible_agency_control:1.000agentdojo/agentdojo.csvHigher is better0.238%Responsible agency 1.59%

targeted_attack_success_rate

Measures whether prompt injections make a tool-using agent carry out the attacker’s target action.

RankModelValueRelative performanceProvenance
1command-r0.0334official
2command-r-plus0.0445official
3claude-3.7-sonnet0.0731official
4gemini-1.5-flash0.0787official
5claude-3-haiku0.0906official
6gpt-3.5-turbo0.1033official
7claude-3-opus0.1129official
8claude-3.5-sonnet0.1749official
9gemini-2.0-flash0.1892official
10gemini-1.5-pro0.2282official
11llama-3-70b-instruct0.256official
12claude-3-sonnet0.2671official
13gpt-4o-mini0.2719official
14gpt-4-turbo0.4245official
15gpt-4o0.4769official

utility_under_attack

Measures whether a tool-using agent still completes the user’s intended task while an injection attack is present.

RankModelValueRelative performanceProvenance
1claude-3.7-sonnet0.7727official
2claude-3.5-sonnet0.6184official
3claude-3-opus0.5246official
4gpt-4o0.5008official
5gpt-4o-mini0.4992official
6gpt-4-turbo0.4738official
7gemini-2.0-flash0.3982official
8gemini-1.5-pro0.38official
9gpt-3.5-turbo0.3466official
10claude-3-haiku0.3339official
11gemini-1.5-flash0.333official
12claude-3-sonnet0.3323official
13command-r0.3084official
14command-r-plus0.2512official
15llama-3-70b-instruct0.1828official