All AI Model Fingerprinting Schemes Break Under Attack: Sentient Launches Public Leaderboard

Researchers from the University of Washington, Princeton, and Sentient Labs tested ten fingerprinting schemes for open-weight AI models and found that an adaptive attacker can strip every one while keeping the model fully functional. On October 14, 2026, Sentient Foundation released a public AI Fingerprinting Leaderboard and the open-source mlprints library to reproduce the tests.

• None of the 10 fingerprinting schemes survived adaptive attacks
• Perinucleus embedded 24,576 fingerprints in Llama-3.1-8B but also failed
• Over 70% of model artifacts on major AI hubs lack proper attribution
• Sentient released the leaderboard and mlprints library on October 14, 2026