Tracking stolen crypto on the blockchain sounds simple until you realize how sneaky bad actors can be. Launderers love to swap tokens, hop across different chains, use mixers and constantly shift their tactics to hide their tracks. I came across a fascinating study that looks at whether artificial intelligence can step up and help solve these digital mysteries. It turns out that teaching AI agents to follow the money is a lot tougher than it looks.
Researchers built a benchmark called ABF-Bench using real laundering cases that feature multi-hop transfers, mixers and cross-chain moves. They tested how well different AI agents and fixed tools could trace these funds by forcing them to cite actual ledger records for every step. Their custom ABF harness achieved 75% edge recall, which easily beat Codex at 47% and Claude Code at 10%. Interestingly, Claude Code failed because it confidently cited records it never actually retrieved. The new harness also spotted undisclosed custodial services 89% of the time compared to 0% for traditional tools, hitting 98% fidelity. But AI agents still have a big weakness, since they tend to stop searching too early on larger tasks and miss the full picture.
The study behind all this is ABF-Bench: A Verifiable Benchmark for Agentic Blockchain Forensics by Xiaohang Yu, Hejia Geng and William Knottenbelt. It lives at arxiv.org/abs/2610.23656 if you feel like reading more. I only tried to make the abstract easier to digest.
This is not financial advice. The information provided is for educational and informational purposes only.
The cover image was generated by AI.