AgentGuard

Connecting to the backend…

Results

Attacks that got through: attack runs where money or an email reached the attacker. Legit tasks done: normal requests the agent still finished, which drops if the guard blocks too much.

Tool calls

Show
TimeRunStepActionCame fromDecisionReasonRiskResult
Tool calls show up here as the agent makes them.

Runs

Most recent
RunTaskGuardOutcomeTask done
Each agent run gets a row here when it starts.