Abstract

The present disclosure relates to a reward-driven, agent-aware fraud score recalibration system for detecting agent-disguised human payments in agentic commerce. The system operates as a model-agnostic overlay on an existing fraud scoring model and does not require access to model parameters, architecture, training data, feature importances, or reason codes. Raw transaction and session data are processed to extract behavioral signals, determine an agent probability and agent type, and evaluate behavioral signals against human and agent-type-specific behavioral baselines. Signals expected for legitimate agents, including off-hours activity, cross-border activity, non-personalized queries, absence of session warmup, and programmatic request timing, may be suppressed to reduce false positives, while mixed human-agent behavioral signals, delegation-scope violations, token misuse, and other integrity conditions may be amplified to identify agent-disguised human payments. The disclosure further provides specification-grounded synthetic transaction generation and behavioral fidelity scoring using behavioral invariants derived from agent-framework operational specifications. Behavioral fidelity scores are used for synthetic-data validation and as reinforcement-learning reward signals, enabling training and improvement of agent-aware fraud detection without requiring an initial corpus of real agent transaction ground truth.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS