Protegrity releases open-weights model for finding personal data in AI workflows

Protegrity has released Protegrity-PII, an open-weights model on Hugging Face that detects and classifies 69 types of personally identifiable information, including card data, health information, credentials, device identifiers and biometric data, Cybersecurity Insiders reports.

The model has 278 million parameters and is built on XLM-RoBERTa. It labels text token by token, treating it as content to analyse rather than instructions to follow, which the company says reduces its exposure to prompt injection. Organisations can run it on their own infrastructure to inspect prompts, tool calls, APIs and agent responses.

Protegrity also published PII-SCOPE, a benchmark of 8,030 records from healthcare, finance, insurance, retail and government, and a technical report on arXiv. On its own benchmark, Protegrity says the model reached 0.910 typed recall and 0.902 precision, ahead of AWS Comprehend, Microsoft Presidio and OpenAI Privacy Filter on both measures, and processed one million characters in 0.49 seconds on a single RTX 6000 Blackwell GPU at an estimated cost of $0.0005.