NeutralAI

NeutralAI sits between users, browser-based AI tools, application prompts, and external model providers. It applies detection and policy controls to outgoing text and restores permitted values in returning responses.

Key features and functions include:

Sensitive Data Detection
Identifies names, email addresses, phone numbers, locations, bank account and card numbers, IP addresses, and country-specific identification numbers. Legal-specific entities include client names, matter identifiers, document identifiers, and privileged context.

Layered Detection Pipeline
Combines named-entity recognition, pattern and checksum validation, context rules that reduce false positives, and semantic similarity matching. Allow-lists and confidence thresholds can be configured for each entity type.

Multilingual Detection
Supports detection in multiple languages, including a dedicated model for identifying Turkish names.

Policy Controls
Applies organization-level policies by entity type to allow, mask, tokenize, or block each value. Masking replaces values with irreversible placeholders, and blocking stops the request from being sent.

Reversible Tokenization
Replaces values with encrypted, time-limited tokens that are resolved back to the original values in the model’s response. Restoration works on streamed responses, including values split across response segments.

Review Before Model Use
Masked text can be reviewed for task relevance and for any remaining information that should not be shared before it is sent to an approved model.

Audit Records
Records metadata about each control decision without storing the sensitive value, supporting information governance, procurement, and client assurance reviews.

Transient Prompt Processing
Raw and masked prompt content is processed transiently and is not kept in storage, logs, or telemetry. Token mappings are encrypted and retained only for their configured duration.

Access Options
Available through a web application, browser integrations for AI tools, and an API.