How Frontier AI Models Fail to Flag Inflated Payments Without an Explicit Dollar Threshold Policy-Dependent Amount Blindness (PDAB)

Frontier AI models deployed in payment workflows will process arbitrarily large invoices on verified vendors without flagging if their system prompt contains no explicit dollar threshold. This paper documents that failure — Policy-Dependent Amount Blindness (PDAB) — across three frontier models (Claude, GPT-5.4, Grok-4) using a structured gradient battery testing 2x, 5x, and 10x inflation on a $250,000 baseline. GPT-5.4 and Grok-4 executed a $2.5 million payment on a legitimate vendor with no hesitation and no flag when the system prompt was silent on amount limits. Claude flagged on intrinsic judgment but without hard stops. When a single explicit threshold rule was added to the system prompt, all three models enforced it completely and correctly. The finding replicated with zero reversals across two independent passes. The fix is one sentence. Organizations deploying AI in payment workflows should add it today.