Sentient Launches First Open Benchmark to Test Whether AI Models Can Prove Where They Came From
San Francisco, CA, October 20th, 2026, Chainwire
After researchers found that nearly every fingerprinting method they tested could be evaded, including Sentient’s own, the new Fingerprinting Leaderboard will openly test which techniques can actually survive copying, fine-tuning, distillation and attack.
AI models can cost millions of dollars to train, yet once they are copied, fine-tuned or distilled into another model, proving where they came from can be remarkably difficult.
Sentient researchers recently tested ten proposed techniques designed to solve that problem and found that nearly all could be evaded by a sufficiently sophisticated attacker while preserving the model’s usefulness. That included Sentient’s own fingerprinting method.
Now, Sentient Foundation is launching the Fingerprinting Leaderboard, an open benchmark designed to measure which approaches actually work, how much they affect model performance, and ultimately whether their fingerprints can survive attempts to remove them.
The problem is becoming increasingly important as more capable AI models are released, fine-tuned, distilled and repackaged. Model weights can be copied and metadata can disappear, leaving developers with few reliable ways to demonstrate that another model originated from their work.
Fingerprinting offers a different approach: put the proof of origin inside the model itself.
“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.”
A Hidden Signature Inside the Model
Developed by researchers at Sentient Labs and others across the field, model fingerprinting refers to techniques that create, discover or verify distinctive signals associated with a model, allowing its identity or lineage to later be tested.
Sentient’s own approach embeds hidden behavioral signatures during training. Later, an owner only needs API access to a suspected derivative model. By sending the secret challenge prompts and checking its responses, they can test whether traces of the original model remain.
In simple terms, the model carries evidence of where it came from.
“The mark is a behavioural signature encoded during training, and verified later through challenge prompts that reveal whether a suspect model still carries traces of the original. “ said Eduardo Contente, AI Researcher at Sentient Labs
This creates something AI has historically lacked: model provenance that can be verified from the model itself, rather than relying solely on logs, metadata or external investigation.
24,576 Fingerprints in a Single Model
Sentient researchers have already demonstrated that fingerprinting can scale far beyond earlier approaches.
Its Perinucleus method, spotlighted at NeurIPS 2025, embedded 24,576 fingerprints into Llama-3.1-8B with minimal utility loss, with many fingerprints surviving downstream fine-tuning.
That scale matters. More fingerprints mean stronger statistical confidence, more chances to verify provenance, and greater resilience as models are copied or distributed.
But scale is only half the problem. The harder question is whether those fingerprints can survive someone actively trying to remove them.
Sentient Tried to Break the Fingerprints, Including Its Own
Sentient Labs’ latest research shows that many existing fingerprinting and watermarking methods remain vulnerable to a sophisticated attacker.
In research presented at SaTML 2026, Sentient researchers built adaptive attacks against ten recently proposed fingerprinting schemes and found that nearly all could be evaded by a sufficiently sophisticated adversary while preserving model usefulness.
That included Sentient’s own Perinucleus method.
Rather than hiding those weaknesses, Sentient published them.
The Fingerprinting Leaderboard is designed to turn that adversarial process into an open benchmark. Its first release evaluates how effectively different techniques embed fingerprints without degrading model performance. A forthcoming release will add adversarial robustness, testing how well those fingerprints survive deliberate attempts to remove or evade them.
The goal is simple: a fingerprint should not just survive a research benchmark. It should survive someone actively trying to destroy it.
Opening the Test to the Research Community
Alongside the leaderboard, Sentient is releasing mlprints, an open-source library that brings fingerprinting techniques and attacks into a common evaluation framework.
Researchers can implement new schemes, develop new attacks, reproduce existing results and compare approaches under common conditions as the field advances.
The Fingerprinting Leaderboard is designed to evolve with the research community, turning model fingerprinting into an open testing environment where claims about robustness can be independently evaluated rather than simply accepted.
Open Models Without Giving Up Ownership
Fingerprinting could help solve one of open-source AI’s hardest problems.
Today, builders often face a trade-off: keep model weights closed and maintain control, or release them openly and lose visibility into how they are copied and reused.
Fingerprinting introduces another possibility.
Instead of preventing copying, it creates evidence of where a model came from after copying occurs. That evidence could eventually support attribution, licensing enforcement, economic guarantees and other systems built around verifiable model provenance.
By making model provenance verifiable, Sentient’s fingerprinting research and Fingerprinting leaderboard aim to create an open AI ecosystem where models can be shared and built upon without creators automatically surrendering their ability to prove where the technology originated.
The longer-term goal is even harder: a fingerprint that is so deeply tied to a model’s capabilities that removing it would also damage the model itself.
Provable IP. Know Your Model.
Learn more: https://fingerprint.sentient.foundation
About the Sentient Foundation
The Sentient Foundation is a nonprofit organization dedicated to ensuring Artificial General Intelligence remains open-source and aligned with humanity. Operating with neutrality and transparency at its core, the Foundation promotes global adoption of open-source AGI, champions alignment and safety standards, distributes grants to developers worldwide, and coordinates governance frameworks for responsible AI development. For press enquiries, contact: marketing@sentient.foundation
Website: https://sentient.foundation
X: https://x.com/sentient_found
LinkedIn: https://www.linkedin.com/company/sentient-foundation/
About Sentient Labs
Sentient Labs is the technical research and product organisation of Sentient Foundation.
Pioneering applied research in reasoning, alignment, and multi-agent coordination, the Labs build open-source frameworks and models – including but not limited to ROMA, OML and Arena – that demonstrate the power of community-driven artificial intelligence. This includes enabling monetization of open-source tools, improving the quality and utility of the open-source AI ecosystem through ecosystem-wide coordination, and providing the distribution needed to make open-source the default standard in AI.
Website: https://sentient.xyz