AWS engineers describe a configurable, instruction-driven PII detector that runs on any LLM served via Amazon Bedrock or self-hosted, shipped as the open-source pii-detector package. It replaces fixed token-classification taggers with prompt-defined entity schemas, so new entity types require a one-line prompt edit instead of retraining. It was benchmarked span-for-span against eight other LLM-based detectors, including OpenAI PrivacyFilter, across five public corpora totaling 49,365 records and 222,114 ground-truth spans in eight languages.
A middleware layer sitting between an application and an external LLM that detects sensitive or personally identifiable information locally and replaces it with consistent placeholders before the request leaves. The LLM still works with the relevant context but never sees the actual PII, letting developers use external models on private data.
A middleware layer sitting between an application and an external LLM that detects sensitive or personally identifiable information locally and replaces it with consistent placeholders before the request leaves. The LLM still works with the relevant context but never sees the actual PII, letting developers use external models on private data.
A tutorial pipeline that redacts personally identifiable information from high volumes of scanned documents like medical forms, insurance claims, and financial records. It uses Amazon Bedrock Data Automation to detect and selectively redact PII where only some instances are sensitive per the use case. Aimed at organizations with compliance-driven document sharing needs where manual redaction is too slow and error-prone.
A tutorial pipeline that redacts personally identifiable information from high volumes of scanned documents like medical forms, insurance claims, and financial records. It uses Amazon Bedrock Data Automation to detect and selectively redact PII where only some instances are sensitive per the use case. Aimed at organizations with compliance-driven document sharing needs where manual redaction is too slow and error-prone.
LocalMask is a command-line tool that runs locally to detect and mask personally identifiable information (PII) and secrets before data is sent to large language models. It solves the privacy and compliance problem for developers and organizations that need to use LLMs without exposing sensitive data.
LocalMask is a command-line tool that runs locally to detect and mask personally identifiable information (PII) and secrets before data is sent to large language models. It solves the privacy and compliance problem for developers and organizations that need to use LLMs without exposing sensitive data.
Prompt-scrub is a local-first tool that redacts personally identifiable information (PII) from LLM prompts and responses before they leave the user's environment. It helps developers and organizations protect sensitive data when using LLM APIs, ensuring privacy and compliance.
Prompt-scrub is a local-first tool that redacts personally identifiable information (PII) from LLM prompts and responses before they leave the user's environment. It helps developers and organizations protect sensitive data when using LLM APIs, ensuring privacy and compliance.
A new arXiv paper, "Mind the Gap: Robustness Risks in PII Detection Systems," shows that modern PII detection systems, despite strong benchmark performance, suffer substantial robustness failures under realistic distribution shifts. The authors argue missed PII entities constitute direct privacy and security risks, making these failures consequential for production deployments. The study analyzes how different PII detection paradigms break down on noisy, unstructured, and informal inputs rather than comparing state-of-the-art accuracy.
A new arXiv paper, "Mind the Gap: Robustness Risks in PII Detection Systems," shows that modern PII detection systems, despite strong benchmark performance, suffer substantial robustness failures under realistic distribution shifts. The authors argue missed PII entities constitute direct privacy and security risks, making these failures consequential for production deployments. The study analyzes how different PII detection paradigms break down on noisy, unstructured, and informal inputs rather than comparing state-of-the-art accuracy.