How PrivacyScrubber Removes PII from AI Prompts

PrivacyScrubber Agentic Guard operates as a local in-memory proxy between your workstation and autonomous AI agents (Cursor, Windsurf, Claude Code, GitHub Copilot). When an agent attempts to execute a shell command (via guard_exec) or inspect configuration files (via guard_read_file), the output is intercepted in volatile RAM, scanned across 30 industry detection profiles, and all API keys, database credentials, and customer PII are replaced with anonymous tokens (e.g. [API_KEY_1], [EMAIL_1]). The cloud LLM reasons only over sanitized placeholders. When the agent writes a patch or updates code (via guard_apply_patch), the tokens are reversed back to authentic local secrets in RAM and written safely to disk with an automatic .bak backup. Zero secrets ever touch cloud servers.

No. Zero bytes of your text, spreadsheet, code, or document data are ever transmitted to any external server. The entire PII detection and tokenization engine runs inside your browser's local JavaScript/Wasm runtime in volatile RAM. You can verify this by opening Chrome DevTools → Network tab and clicking "Scrub in RAM (0 Bytes Egress)" — exactly 0 outbound network requests will occur.

Because PrivacyScrubber is architected with Zero Servers processing prompt data, there is no centralized user database, no credential storage, and no password hashing infrastructure. Operating without user accounts eliminates account credential theft and credential stuffing as attack vectors. Furthermore, you don't need a time-limited free trial: the Community Tier is permanently free (15,000 characters per scrub for universal PII and 5,000 characters for specialized profiles) with zero sign-up friction. Commercial tiers (PRO and TEAMS) activate via client-side cryptographic license keys verified in RAM without requiring an account login.

The session map is stored exclusively in volatile browser tab RAM. The moment you close the tab, reload the page, or navigate away, the entire mapping — including all original PII values — is permanently destroyed. It is never written to localStorage, cookies, IndexedDB, or disk storage.

AI Tokens mode replaces sensitive entities with structured, context-preserving badges (e.g. [NAME_1], [EMAIL_1]). This allows AI models (ChatGPT, Claude, Gemini) to comprehend entity roles and logical flow while preventing personal data ingestion, and enables 1-click original recovery via Reveal. Blackout mode is designed for court filings, legal discovery, and FOIA disclosures: it renders permanent, solid black raster bars over pixel coordinates and completely purges the underlying vector text layer to ensure zero copy-paste leakage.

Most conventional PDF viewers, annotation apps, and superficial masking tools merely draw black vector rectangles on top of the visual layer. Underneath those black boxes, the raw OCR text stream, character coordinates, and hidden XML metadata remain 100% intact — allowing anyone to copy the redacted text via Cmd+A → Copy, or extract it with command-line tools like pdftotext. PrivacyScrubber executes true Secure Raster Flattening: each page is rendered onto an off-screen graphics canvas in local RAM, PII coordinates are overwritten at the pixel level (either with solid blackout bars or contextual AI Tokens), and a brand-new PDF is synthesized from clean raster bitmaps. The original text layer is physically destroyed and cannot be recovered.

No. Because PrivacyScrubber never receives, processes, or stores your data on its servers, PrivacyScrubber is not acting as a HIPAA Business Associate or a GDPR Data Processor. Organizations typically determine that standard Business Associate Agreements (BAAs) and Data Processing Agreements (DPAs) are not applicable, eliminating 3–6 month procurement security audit delays.

The CISO Procurement Memo is a standardized, copy-ready compliance document for enterprise IT, security architects, and procurement officers. It provides General Counsel, DPOs, and InfoSec committees with formal statutory evidence (GDPR Art. 4(8)/28, HIPAA 45 CFR § 164.502(e), CCPA § 1798.140) proving that because PrivacyScrubber runs 100% in local client RAM with 0 bytes transmitted to any server, third-party DPA/BAA contracts and lengthy vendor risk questionnaires (SIG/CAIQ) are legally non-applicable, enabling fast-track 0-day software approval.

The Free tier includes up to 15,000 characters per scrub for the General PII profile (names, emails, phones, IDs, addresses, IPs) and a 5,000-character free quota across all 30 specialized profiles, with single .txt/.docx file upload. The PRO plan ($15/mo or $110 Lifetime) provides unlimited character processing across all 30 industry profiles (Healthcare PHI, Financial IBAN/Tax, Legal NDAs), cell-by-cell Excel spreadsheet scrub (.xlsx/.csv), offline PDF/OCR scanning, batch document processing, and unlimited Custom Regex rules.

Using an in-memory SheetJS Wasm engine, PrivacyScrubber parses workbook cells individually in local RAM. It identifies sensitive names, emails, and financial metrics within table cells and replaces them with standardized tokens while preserving sheet headers, column boundaries, and mathematical formulas for accurate analysis in ChatGPT Advanced Data Analysis.

When you highlight any text on any webpage or web application, the extension spawns a small floating shield icon (.ps-floating-shield). Clicking it immediately tokenizes the selected text in-place and opens a clean popover showing all identified entity chips with 1-click false-positive unmasking.

Yes. PrivacyScrubber offers an official local Model Context Protocol server package @privacyscrubber/mcp-server that integrates directly with Cursor IDE, Windsurf, Cline, and Claude Desktop. It acts as an offline pre-flight prompt security gateway, intercepting API keys, connection strings, and PII locally in RAM before they leave your workstation.

Under the TEAMS subscription ($99/mo flat), an administrator encrypts their active tab session map using Argon2id key derivation and XChaCha20-Poly1305 authenticated encryption. This generates an encrypted link or file payload. Colleagues decrypt the token map locally on their machines using the shared passphrase — no database or cloud service ever touches your keys.

The purchaser acts as the Team Administrator with exclusive authority to configure organization-wide custom Regex rules, corporate token taxonomy (e.g. [PATIENT_ID]), and default compliance profiles in the Teams Dashboard. When the Admin exports a cryptographically signed Blueprint with Lock Rules enabled, Managed Team Members receive a locked, read-only configuration across their Web App and Chrome Extension (badged as Enterprise), preventing employees from altering or bypassing corporate DLP policies.

PrivacyScrubber encloses both the editor input (#inputText) and highlighted token output within isolated, closed Shadow DOM roots. This makes the raw, unredacted text completely inaccessible to any third-party scripts, browser extensions, or key-sniffing scripts running in your main document pool, creating a hardened frontend sandbox.

To prevent PII or documents from lingering in device memory, the processing engine actively overwrites temporary parsing buffers (used for DOCX, PDF, and spreadsheet extraction) at the byte level using Uint8Array.fill(0) in local memory immediately after use, ensuring zero sensitive residue remains in RAM before garbage collection.

Custom Regex Rules are a pattern discovery engine used to identify proprietary identifiers (such as internal ticket numbers, case IDs, or custom account formats). In contrast, Token Labels rename standard placeholders (for example, mapping NAME → PATIENT or ID → CASE_NUMBER) so that the LLM receives domain-accurate semantic context in its prompt without custom regex code.

Yes. If your procurement department requires manual invoicing, purchase orders (PO), or bank wire transfers (SWIFT/SEPA), our team issues formal commercial invoices with VAT/tax reverse-charge support. Contact support@privacyscrubber.com to set up manual billing.

Because PrivacyScrubber maintains a Zero-Trust Data Sanitization (ZTDS) architecture with zero server databases and zero phone-home user tracking, license custody is sovereign to the purchaser. All license tiers are strictly non-transferable. If a developer accidentally commits a key to a public repository, or an employee misplaces their credentials, the billing owner can immediately request emergency revocation via support@privacyscrubber.com. We append the SHA-256 hash of the compromised key to our global Key Revocation List (KRL), instantly invalidating it across all client instances within 24 hours while reissuing a secure replacement.

When you drag and drop or upload a document (such as a resume or agreement) into ChatGPT or Claude, PrivacyScrubber intercepts the file in local browser RAM before it transmits across the network. Binary formats (.pdf, .docx) inherently contain embedded metadata, hidden XML streams, author revisions, and system properties that leak identity even if visual text is altered. PrivacyScrubber extracts the text locally, scrubs all PII tokens, strips identity from the filename itself, and injects a clean synthetic .txt payload into the AI input. This eliminates 100% of hidden binary metadata while allowing LLMs to parse the content natively. The Free tier includes a daily quota of 3 document uploads per 24 hours (and up to 15,000 characters per scrub), while the PRO and TEAMS plans provide unlimited document processing and custom rules.

Each license tier is scoped to specific operational boundaries:

• PRO ($15/mo or $110 Lifetime): Single-user personal license valid on up to 3 personal devices (Web, Chrome, IDE MCP). Strictly non-transferable.
• TEAMS ($99/mo flat): Organization-wide internal workspace covering unlimited internal staff with Master Key Isolation. Staff access the tool via encrypted Magic Links or managed Chrome policies without exposing master credentials.
• DEVELOPER SDK ($299/mo or $2,990/yr): Headless execution across unlimited internal backend nodes, microservices, ETL pipelines, and internal RAG vector databases. Embedding into customer-facing commercial SaaS for third-party resale requires an OEM/Enterprise contract.

Developers integrate @privacyscrubber/sdk with zero dependencies using a single-line client wrapper like wrapOpenAI(new OpenAI()), or via standalone sanitize() and restore() in under 1ms local RAM latency. It intercepts outbound prompts before network transmission, tokenizes PII and infrastructure secrets into semantic placeholders ([NAME_1], [AWS_KEY_1]), and rehydrates LLM responses locally—including streaming SSE tokens via sliding window reconstruction. Operation is completely plug-and-play with zero background servers or Python/spaCy daemons. The primary operational risk to manage is state custody in distributed microservices: tokenMap is volatile and RAM-only, requiring stateless microservices to preserve or pass the token map between pre-processing and post-processing nodes.

Yes. The @privacyscrubber/mcp-server package executes strictly in local machine memory over standard stdio JSON-RPC transport. It makes zero outbound network requests, transmits zero telemetry, and performs all PII sanitization and token mapping exclusively in your local volatile RAM, satisfying stringent IDE security policies across Cursor, Windsurf, and Claude Desktop.

Every software build undergoes an automated 13-suite security and performance QA pipeline in continuous integration (CI). Automated headless network sniffers verify that during PII sanitization, exactly 0 HTTP, WebSocket, or background telemetry packets leave the browser. In-memory heap profiling enforces an ultra-lightweight ceiling under 6MB RAM (5.68 MB JS Heap) and sub-75ms streaming DOM mutation scans across 5,000 live nodes, ensuring zero laptop heating and zero CPU throttle. Review the full audit in our Technical Security Specifications.

No. When you drop or upload a document (.txt, .docx, .pdf, or client meeting transcript) into PrivacyScrubber, the file never travels across the internet. The entire webpage operates like a self-contained local application inside your browser: file parsing, text extraction, and PII redaction execute exclusively in your computer's local RAM. You can verify this complete data isolation by turning your computer's Wi-Fi completely off (Airplane Mode) before dropping your file — the tool continues to sanitize your documents with 100% functionality and zero network connectivity.

Solo practitioners handling confidential client sessions (coaching, therapy, executive consulting) can execute a 100% offline, zero-cloud workflow: First, transcribe your locally recorded Zoom or Teams meeting on your own device using an offline Speech-to-Text engine (such as MacWhisper or Whisper.cpp running natively on Apple Silicon / local GPU). Second, drop the resulting .txt or .docx transcript into PrivacyScrubber at privacyscrubber.com (even with your device Wi-Fi completely turned off). PrivacyScrubber strips client names, sensitive disclosures, neurodiversity/medical conditions, and identifying details into safe tokens ([NAME_1], [ORGANIZATION_1]) in volatile RAM. Third, submit the sanitized transcript to Claude or ChatGPT to extract themes, check coaching quality, or draft session summaries. Finally, click Reveal in PrivacyScrubber to rehydrate original client names locally. Zero client personal data ever touches third-party servers.

PrivacyScrubber features dedicated pattern recognition engines for European business and civil nomenclature. Our punctuation-aware lookahead models accurately tokenize initial-prefixed names with noble particles (e.g., 'J. de Ruiter', 'K. van den Berg', 'M. de Jong') without prematurely splitting or dropping initials. For statutory identity tokens, PrivacyScrubber embeds native checksum and format validators in client-side RAM for European national IDs—including the Dutch 11-test checksum for BSN/SOFI numbers, 11-digit German Steuer-ID, 16-character Italian Codice Fiscale, Spanish DNI/NIE, French NIR/INSEE, and Dutch KvK commercial register numbers.

Embedding raw personal data into vector databases (Pinecone, Qdrant, Weaviate, Milvus) creates permanent regulatory liability: fulfilling a single GDPR Article 17 'Right to Erasure' request requires locating all high-dimensional embeddings containing that person's data and recalculating vector indexes—costing upwards of $14,000 per deletion cycle. By embedding the headless @privacyscrubber/sdk into your ingestion pipeline, raw PII is swapped with deterministic cryptographic tokens in Node.js heap (<0.065ms) before vector embedding. When a deletion request arrives, organizations simply purge the local key mapping, achieving instantaneous O(1) mathematical erasure at $0 infrastructure cost.