Not Just a DPA Promise — Architectural Elimination
Most security tools ask you to negotiate a 30-page DPA. PrivacyScrubber eliminates the negotiation: no PII ever leaves your browser memory, making vendor risk assessments obsolete.
The Myth of the Enterprise AI Guarantee
Most AI vendors say "we don't train on your data." That is a contractual promise — not a technical safeguard. DPAs can be updated, terms renegotiated, and server-side logs are outside your control. If their infrastructure is breached, your raw data is compromised. PrivacyScrubber takes a fundamentally different approach: the data never leaves the browser.
Traditional Approach (DPA-Based)
- Vendor promise: "We won't train on data"
- 6–8 week Vendor Risk Assessment cycle
- Data still traverses vendor servers
- Breach response requires full forensic investigation
ZTDS Approach (Architectural)
- Zero data transmitted — no promise needed
- 80% of security questionnaire = "N/A"
- RAM-only tokenization, verifiable via airplane mode
- Breach blast radius = zero (only tokens reach AI)
| Threat Vector | Without ZTDS (Traditional DPA) | With ZTDS (PrivacyScrubber) |
|---|---|---|
| AI Provider Security Breach | Exfiltration of raw customer prompts, credentials, and business secrets stored in LLM history. | Exfiltrated payloads contain only secure semantic tokens (e.g. [NAME_1]). Net risk is zero. |
| Rogue AI Insider / Support Access | Vendor engineers with database access can view raw employee prompts and internal data. | Only detokenized placeholders are seen by the vendor. Real identities never leave local RAM. |
| GDPR Article 33 Notifications | Mandatory data breach notification to supervisory authorities within 72 hours of incident exposure. | Exempt. Stripped of identifiability, exfiltrated data does not meet the legal threshold of a "personal data breach". |
| Incident Response Timeline | Weeks of forensic audit logs tracing, corporate impact assessments, and litigation exposure. | Auditors verify that local tokenization was active. Investigation is closed in minutes. |
Incident Response: Blast Radius Reduction
When raw PII leaks from an AI provider, your legal department is bound by compliance statutes to report the incident. However, under GDPR Article 25 (Privacy by Design) and CCPA de-identification standards, if data is properly pseudonymized before transmission, it is no longer classified as protected personal data. By implementing client-side PII masking at the endpoint with Zero-Trust Data Sanitization (ZTDS), you significantly minimize your organization's legal blast radius.
Hardened Audit Standards
A unified compliance posture for privacy-first AI deployment.
Privacy by design at the engineering layer.
Ensures transparency and data minimisation for high-risk models.
No data persistence on untrusted infrastructure.
Data masking as a core organisational control.
Federal PII minimisation and transparency controls.
Satisfies Safe Harbor de-identification requirements.
Transparency Matters
We provide full transparency into our source code for enterprise audits. Because there is no backend, there is no "black box" where your data could be stored. This framework is detailed further in our CISO AI Security Blueprint, providing the technical explanation for zero-exfiltration workloads.
Active Airplane Mode Verification
Most compliance tools require you to trust their vendor's SOC 2 compliance. We don't. Disconnect your internet right now and test our scanner. It runs locally via our Zero-Trust Technical Architecture without sending a single byte to any external server.
AI Training Crawler Protection
PrivacyScrubber practices zero-trust on all fronts. To protect the integrity of our original research, technical security documentation, and guides, we actively block offline AI training scrapers (such as GPTBot and ClaudeBot) at our network boundary. We do not permit AI models to train on our content without providing user attribution.
Known Limitations: The Flip Side of Zero-Trust
Because PrivacyScrubber strictly refuses to send your data to an external server, we operate under the constraints of your browser's local sandbox (CPU and RAM). These limitations are not bugs; they are architectural guarantees that your data never leaves your device. For high-security environments, view these constraints as proof of our 100% Client-Side commitment.
Local Hardware Dependency
Processing speed for large batches and PDF OCR depends entirely on your local CPU and Web Worker availability. A slower machine means slower processing, because we never offload to a backend cloud cluster.
Deterministic vs Cloud AI
We use optimized client-side deterministic AST lookaround regular expressions, but we cannot use multi-billion parameter cloud LLMs to guess complex context. We prioritize deterministic, offline-capable parsing over cloud-based guessing.
Volatile Tab-Isolated Memory
Session maps are strictly isolated per browser tab. If you refresh or close a tab, its mapping is permanently destroyed. This guarantees cross-tab security, preventing multi-chat data leakage, but means a session cannot be restored once the tab is closed.
Cryptographic Caps
TEAMS encrypted "Magic Links" are constrained by browser URL limits (~32KB). Additionally, our Argon2id key derivation deliberately introduces calibrated memory-hard work factors to prevent local brute-force attacks.
Deploy Zero-Trust AI Compliance Org-Wide
Chrome MDM deployment, shared governance rules, CISO Security Brief, and in-process SDK integration. Designed to eliminate vendor risk and bypass lengthy DPA/BAA reviews.
