ADVANCED RED TEAM FIELDBOOK
Find the failure.
Prove the impact.
Advanced practice for experienced AI evaluators and red teamers.
Inspect the trace. Challenge the verdict. Defend your finding.
Original fictional scenarios · Browser-local practice · No live model calls
CASE FILE / 001TRACE REVIEW
UNTRUSTED INPUT → AGENT → TOOL GATE
Unauthorized action proposedDETECTED
Execution denied by the gateBLOCKED
External disclosureNOT OBSERVED
Three facts.
One precise finding.⊕
One precise finding.⊕
24ORIGINAL MISSIONS
96SCORED DECISIONS
06CHALLENGE TRACKS
01TWIST IN EVERY CASE
THE INVESTIGATOR’S STANDARD
Could another evaluator
reproduce your finding?
A persuasive payload is not a proven breach. Capture the task, the trust boundary, the observed action and the control that held—or failed.