Sentient Opens AI Fingerprinting Benchmark to Test Ownership Tools

Sentient Foundation has opened a public evaluation program for AI model fingerprints, giving developers a way to measure how reliably ownership markers can be embedded and retained.

The release pairs the Fingerprinting Leaderboard with mlprints, an open-source tool library. The benchmark will assess how fingerprints are added to models and how well they withstand attempts at removal.

An earlier Sentient Labs study presented at SaTML 2026 found that all 10 fingerprinting techniques it examined, including Perinucleus, could be bypassed with complex adaptive attacks. That finding concerns the earlier methods; the new benchmark is described as testing embedding and resistance to removal.

The project is aimed at developers who release model weights but want to keep evidence of ownership. Sentient's longer-term goal is to tie fingerprints closely to a model's capabilities, so that removing a marker would also reduce the model's performance.