Agents
FinHarness: Inline Lifecycle Safety Harness Cuts Finance LLM Agent Attack Success Rate from 38% to 15%
Published May 26 on arXiv (2605.27333), FinHarness introduces a three-component safety system — Query Monitor, Tool Monitor, and Cascade module — that monitors financial agents throughout their entire operational lifecycle rather than only at entry or exit points. On the FinVault benchmark, the approach reduced Attack Success Rate from 38.3% to 15.0% while maintaining 39.3% benign approval (down only 1.8% from 41.1%) and using 4.7x fewer advanced-judge calls compared to always-on verification. The key insight is that mid-process tool calls in multi-step financial workflows represent a critical blind spot that boundary-only filters miss entirely.
Source
↳ Follow the thread