Sentient Launches First Open Benchmark to Test Whether AI Models Can Prove Their Origins

Sentient Foundation has launched a new open benchmark to evaluate AI model fingerprinting techniques, following research showing that nearly all of the methods tested, including its own, could be evaded by sophisticated attackers.

As AI models become more valuable and widely distributed, determining whether one model was copied or derived from another is emerging as a significant challenge for developers. Models can cost millions of dollars to train, yet once their weights are copied, fine-tuned or distilled, establishing their origin can become difficult.

Sentient Foundation is seeking to address this problem with the launch of its Fingerprinting Leaderboard, an open benchmark designed to evaluate how effectively different techniques can establish model provenance without compromising performance.

The initiative follows research by Sentient Labs that examined ten recently proposed fingerprinting schemes and found that nearly all were vulnerable to adaptive attacks, including Sentient’s own Perinucleus method.

Why AI Model Fingerprinting Matters

Traditional methods of tracking model ownership often rely on metadata, logs or external records. However, these can become unavailable when models are copied, modified or redistributed.

AI model fingerprinting takes a different approach by creating or identifying distinctive signals within a model that can later be used to verify its identity or lineage.

Sentient’s approach involves embedding hidden behavioral signatures during training. Model owners can subsequently test a suspected derivative through API access, using secret challenge prompts to determine whether traces of the original model remain.

“AI models are becoming some of the most valuable intellectual property in the world, yet once a model is copied or distilled, proving ownership becomes incredibly difficult,” said Sachi Kamiya, Director of Venture and Growth at Sentient Foundation. “Fingerprinting puts the proof inside the model. Think of it as the AI equivalent of the hidden marks Renaissance painters used to establish authorship.”

The technology could provide developers with a way to establish model provenance even after their models have been modified or distributed.

Research Reveals Limitations in Existing Techniques

Sentient Labs has previously demonstrated the potential scale of model fingerprinting through its Perinucleus method, which was spotlighted at NeurIPS 2025.

Researchers embedded 24,576 fingerprints into Llama-3.1-8B with minimal impact on model utility, with many of the fingerprints surviving subsequent fine-tuning.

However, the ability to embed fingerprints does not necessarily mean they can withstand deliberate removal.

In research presented at SaTML 2026, Sentient researchers developed adaptive attacks against ten recently proposed fingerprinting schemes. Their findings showed that nearly all could be evaded by a sufficiently sophisticated adversary while preserving the model’s usefulness.

The results also exposed vulnerabilities in Sentient’s own Perinucleus technique, highlighting the gap between demonstrating fingerprinting capabilities and proving their resilience against attacks.

A Public Benchmark for Testing AI Model Provenance

The newly launched Fingerprinting Leaderboard aims to establish a common evaluation framework for researchers developing and testing fingerprinting methods.

Its initial release focuses on measuring how effectively different approaches embed fingerprints while maintaining model performance.

A forthcoming release will introduce adversarial robustness testing, examining whether fingerprints can survive deliberate attempts to remove or evade them.

Alongside the leaderboard, Sentient is releasing mlprints, an open-source library that brings fingerprinting techniques and attacks into a shared evaluation environment.

The library allows researchers to implement new methods, reproduce existing results, develop attacks and compare techniques under common conditions.

By making these evaluations accessible to the research community, Sentient aims to enable independent testing of fingerprinting claims rather than relying solely on results published by individual developers.

Could Fingerprinting Help Protect Open-Source AI?

The development also addresses a broader challenge facing open-source AI developers: how to make models publicly available without losing the ability to demonstrate their origin.

Developers currently face a trade-off between keeping model weights private to maintain control and releasing them openly, where subsequent copying and modification can be difficult to track.

Fingerprinting could offer an alternative by preserving evidence of a model’s origins even after it has been copied or adapted.

If sufficiently robust, such techniques could eventually support attribution, licensing enforcement and economic guarantees for developers distributing open models.

However, Sentient’s own research suggests that significant technical challenges remain, particularly when attackers deliberately attempt to remove identifying signals without damaging a model’s capabilities.

The longer-term objective is to develop fingerprints so closely connected to a model’s capabilities that removing them would also reduce its usefulness.

With the launch of the Fingerprinting Leaderboard and mlprints, Sentient is opening that challenge to the wider research community, providing a framework for evaluating which methods can withstand increasingly sophisticated attempts at evasion.

The Fingerprinting Leaderboard is available at https://fingerprint.sentient.foundation.