DeXposure-Claw: New AI System for DeFi Risk Supervision and Fraud Detection

Decentralized finance (DeFi) presents unique challenges for risk supervision due to its rapid evolution and interconnected credit risks. Traditional general-purpose Large Language Model (LLM) agents are often ill-suited for this environment, tending to over-interpret weak evidence and recommend high-stakes interventions, leading to an unacceptable rate of false alarms. This research addresses these issues by introducing DeXposure-Claw, a specialized agentic supervision system. DeXposure-Claw operates on a forecast-grounded approach, routing LLM decisions through structured evidence. Its core components include DeXposure-FM, a graph time-series foundation model that predicts future exposure networks. These forecasts are then fed into deterministic monitors and stress scenarios, which generate specific alerts, attribution signals, and scenario-based evidence. Before any escalation, the system incorporates data-health and confidence gates to constrain interventions, ensuring that only well-supported supervisory tickets with clear rationales are emitted. To validate its effectiveness, the authors developed DeXposure-Bench, a six-axis evaluation framework that measures the system's performance against regulator-aligned absolute-loss ground truth and explicitly tracks false intervention rates. Experiments using five years of real-world weekly data fully support the system's capabilities.