Blockchain forensics preprint reports stronger evidence tracing by AI agents

Researchers introduced ABF-Bench, a benchmark for tracing blockchain funds, and reported that their ABF agent harness recovered 75% of transaction links, compared with 47% for Codex using the same model and 10% for Claude Code. Each claimed movement must cite ledger records retrieved by the agent. The preprint builds tasks from real laundering activity, including exchanges, cross-chain transfers and mixer deposits. The harness recognized undisclosed custodial services at 89%, versus 0% for fixed-order tracers, with 98% fidelity. However, agents stopped early and missed more links on larger tasks, leaving their final recall below the best fixed-order tracer.