marijastavric24/safeprompt

SafePrompt is a browser extension that scans and redacts sensitive text on the user's own device before it is submitted to an AI chat tool like Claude.ai, ChatGPT, Gemini, or Microsoft Copilot. The README repeats that all detection runs locally and that no text is sent to a server. The intended audience is people who work with confidential information: lawyers, HR teams, finance staff, healthcare workers, journalists, and anyone subject to rules like GDPR or HIPAA. A Redact button is injected into the chat input on a supported site. Clicking it scans the text with compiled regex rules and a bundled NLP library, highlights what looks sensitive, and lets the user approve or remove each match. Clicking Apply replaces the matched spans with a [REDACTED] label before submission. Supported browsers are Chrome 112 or newer, Edge, and Firefox 115 or newer. The detection list covers general PII such as email addresses, phone numbers, dates of birth, home addresses, UK postcodes, US SSN and UK NI numbers, passport numbers, Luhn-validated cards, mod-97 validated IBANs with sort codes, IP addresses, API keys, and bearer tokens. Domain rules add legal context like client names, case numbers, and settlement figures, HR items such as salaries, healthcare data like patient names and NHS numbers, financial identifiers (UTR, VAT, SWIFT or BIC), and journalism cues near phrases like 'source:'. Custom rules can be added from a Prompt Library. Each match gets a confidence score. Detections at 85 to 97 percent are auto-approved, 60 to 84 percent are flagged for manual review, and below 60 percent are blocked until acknowledged. The document score is capped at 92 percent and no single match reaches 100 percent. The default is to redact everything: the review panel asks the user what to keep, and personal names, phone numbers, and direct financial identifiers cannot be kept. The security section claims no network calls, enforced by a content security policy of connect-src 'none' that can be verified in browser DevTools, plus local-only detection, in-memory handling of the original text, and minimal permissions (storage and activeTab only). The build uses Node.js 18 or newer with esbuild. The source tree contains content scripts, the detection engine, a Shadow DOM panel, an options page, a popup, a learning module, a sandbox for testing rule changes, and shared utilities. A closing disclaimer notes that automated detection makes mistakes and every highlight should still be reviewed.