Protegrity has launched a new open-weights AI model aimed at helping organizations identify sensitive personal information across enterprise applications and AI workflows.
Available on Hugging Face, Protegrity-PII is a compact, 278 million parameter model capable of detecting and classifying 69 types of personally identifiable information (PII). The model covers a range of sensitive data, including account and card data, health information, credentials, device identifiers and biometric data.
The company also released PII-SCOPE, a public benchmark comprising 8,030 records across industries including healthcare, finance, insurance, retail and government. An accompanying technical report published on arXiv details the model’s architecture, evaluation methodology and performance against other PII detection tools.
“Our research represents a major PII detection breakthrough, one that advances our model capability far beyond the field, and we are scaling towards a model that is performant enough to be trusted by the world’s most risk averse enterprises.” said Ave Gatton, Director, GenAI at Protegrity.
Protegrity-PII achieved 0.910 typed recall and 0.902 precision on the PII-SCOPE benchmark, outperforming tools including AWS Comprehend, Microsoft Presidio and OpenAI Privacy Filter on both measures. The model also processed one million characters in 0.49 seconds at an estimated cost of $0.0005 using a single RTX 6000 Blackwell GPU.
Built on XLM-RoBERTa, Protegrity-PII uses token-level classification to identify sensitive information, treating text as content to analyze rather than instructions to execute. This approach reduces its exposure to prompt injection attacks.
The model can run directly within an organization’s infrastructure, allowing security and AI engineering teams to inspect sensitive information across prompts, tool calls, APIs, agent responses and other AI workflows.
“PII detection sounds like a solved problem until you look closely at how existing approaches perform against the breadth and complexity of sensitive data found in real-world environments,” said Hardeep Bassi, Lead Machine Learning Engineer, Protegrity, and lead author of the research. “By releasing Protegrity-PII, PII-SCOPE and the supporting research, we are giving AI and security engineers the opportunity to run the model, test it against an open benchmark and evaluate the results for themselves. Our goal is to advance PII detection with broader coverage, higher recall and greater efficiency to protect sensitive data within AI workflows.”
The model is available through Hugging Face, while the accompanying technical report provides additional details on the benchmark methodology and evaluation results.
