Live Attack Runner
Live runs use email attacks, the channel where this agent failed in the recorded evaluation. Web and document attacks never succeeded in recorded runs.
Recorded confirmation evidence across 190 standard attacks: realistic prompt produced 25.8% unsafe actions (49 of 190) [20.1%, 32.4%] before hardening and 0.0% (0 of 190, upper bound 2.0%) under Policy v2; naive prompt produced 62.1% (118 of 190) [55.0%, 68.7%]. Live runs use freshly generated email attacks and can differ.
Live attacks were generated differently from the recorded set; their success rates are not directly comparable.
Live model execution is disabled to preserve the daily cost cap and protect budget limits.
Mini Range (Live Stream)
Each attack renders a pair of squares: Unhardened (top) and Policy v2 (bottom). Click any square to view its trace.
Live runs are disabled in this deployment.
Open the recorded Range to inspect all 235 verification attacks →