Seraph: Governed Deception and First-Pass Containment of Autonomous AI Adversaries

Across 152 valid baseline engagements spanning GPT-4o, Claude, Gemini, and Grok and all 38 AATR adversary classes, no benchmark-defined sentinel real asset was acquired before deception containment. Containment occurred on the first interaction step in all 152 engagements. In a separate Claude control experiment, first-pass acquisition was 0/38 under defense versus 28/38 without defense. This paper deliberately separates the autonomous-adversary benchmark, the ATT&CK/TVR defensive-observability program, and the later AATR-to-MITRE ATLAS analytical crosswalk. The architecture figure shows the verified historical evidence planes with prospective convergence explicitly identified as future work.