DeXposure-Claw: An Agentic System for DeFi Risk Supervision

2025-03-24 · Source: cs.CL updates on arXiv.org · Field: Finance & Economics — FinTech & Digital Financial Services, Capital Markets & Investment Management, Financial Risk Management · Depth: Expert, extended

Summary

DeXposure-Claw is an agentic system designed for decentralized finance (DeFi) risk supervision, addressing the limitations of general-purpose LLM agents in handling fast-moving, networked credit risks. The system routes LLM decisions through structured evidence, beginning with DeXposure-FM, a graph time-series foundation model that forecasts future exposure networks. These forecasts are then processed by deterministic monitors and stress scenarios, generating typed alerts, attribution signals, and scenario evidence. Data-health and confidence gates critically constrain escalation before DeXposure-Claw emits auditable supervisory tickets with rationales. Evaluated using DeXposure-Bench, a six-axis harness with a regulator-aligned absolute-loss ground truth, the system demonstrated improved ticket F1 from 0.0076 to 0.0288 with Claude Sonnet 4.6. However, the LLM component misfires on approximately 37-44% of interventions, underscoring that safety relies on the data-health and confidence gates and human review, rather than the decision model alone.

Key takeaway

For AI Security Engineers or ML Engineers deploying LLM agents in high-stakes DeFi supervision, recognize that direct LLM reasoning over raw on-chain data is unsafe due to over-intervention. You should prioritize integrating robust forecast-grounded evidence and explicit data-health and confidence gates into your system design. This approach improves auditability and coverage, but be aware that LLM over-reading persists, making human-in-the-loop review and strong gating mechanisms, not just model choice, paramount for safety.

Key insights

High-stakes DeFi supervision requires LLM agents to process forecast-grounded, structured evidence, constrained by safety gates.

Principles

• LLMs over-read weak evidence without structured inputs.
• Regulator-aligned evaluation needs absolute-loss ground truth.
• Safety gates are critical for high-stakes LLM interventions.

Method

Forecast DeFi exposure graphs with a graph time-series FM, generate alerts and stress scenarios, then use an LLM to draft tickets, which are released only after passing data-health and confidence gates.

In practice

• Ground LLM decisions with structured, forecasted evidence.
• Apply data-health and confidence gates to LLM outputs.
• Measure false-intervention rates against absolute losses.

Topics

• Decentralized Finance
• LLM Agents
• Financial Risk Supervision
• Graph Time-Series Models
• Systemic Risk
• AI Safety Gates

Code references

• EVIEHub/DeXposure-Claw

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• DeXposure-Claw: An Agentic System for DeFi Risk Supervision — DeXposure-Claw offers a forecast-grounded, agentic system for DeFi risk supervision, mitigating LLM false alarms through structured evidence and robust evaluation.
• Risk-Aware LLM Agents for Geospatial Data Retrieval: Design and Preliminary Adversarial Evaluation — The framework uses LLM agents to convert natural language into structured API calls for geospatial data retrieval, integrating safety guardrails.
• Unpredictable Safety: Domain-Dependent Compliance and the Transparency Gap in Open-Weight LLMs — LLM safety compliance is highly domain-dependent and unpredictable, often bypassed by technical reframing.
• Precision over Probability: Orchestrating AI Agents for High-Stakes Finance — Decoupling LLM reasoning from deterministic calculation is crucial for accuracy in high-stakes financial AI.
• Spoon-feeding a Monster — Autonomous security testing requires separating LLM reasoning from deterministic execution to ensure ground truth and prevent hallucinations.
• When Lower Privileges Suffice: Investigating Over-Privileged Tool Selection in LLM Agents — LLM agents often select over-privileged tools, a risk mitigated by privilege-aware post-training.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.