Data Protection DFKI Releases Protective Layer for AI Prompts Directly in the Browser

Related Vendor

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

More and more confidential information ends up in prompts for AI chat services today. With Privacy Guardrail, DFKI releases an open-source extension for Google Chrome that detects, anonymizes, and restores personal and other sensitive content entirely locally in the browser after the AI response.

Generative AI has long become part of everyday work—in emails, support tickets, protocols, research notes, or internal documents. However, this is precisely where a new security gap arises: anyone wanting to use AI productively often has to process texts that ideally should not be shared with external services. Privacy Guardrail addresses this exact moment, where confidential text turns into an AI prompt.

Data Protection at the Point of Origin

The extension automatically checks inserted content locally, marks detected sensitive areas, and replaces them with typed placeholders like [EMAIL_1] or [PERSON_1] before sending. Users can review, adjust, or intentionally ignore individual detections before submission. After the AI system's response, known placeholders can be locally replaced with the original values, preserving the contextual meaning. The key difference compared to many other approaches lies in the local-first approach: detection, placeholder assignment, anonymization, and restoration are all carried out entirely within the browser. Inserted texts are not sent to an external inference service. Privacy Guardrail thus turns the browser itself into the place of data protection—directly where sensitive content becomes a prompt. All relevant data remains in the local Chrome profile of the respective browser. This includes, depending on usage, settings, placeholder mappings, identity vault entries, as well as local correction and feedback data; no storage occurs in Chrome Sync. This ensures control over sensitive information remains with the users.

Two Local Detection Levels

Technically, Privacy Guardrail combines two local detection layers. Deterministic pattern recognizers detect structured content such as email addresses, credit card numbers, IBANs, or IP addresses. Additionally, the local AI component can recognize context-dependent terms such as people, organizations, addresses, places, or passwords. If WebGPU is available, inference runs locally via the graphics card; otherwise, the system uses a slower CPU/WASM path. For devices with limited resources, the extension can also switch to a pattern-only mode, where structured formats are still detected, but the coverage of free text areas is reduced.

Limitations and Transparency

For so-called low-signal categories like URL, DATE, or MISC, detection performance is limited. The DFKI communicates these limitations openly: sensitive content may be overlooked, harmless content may be mistakenly deemed critical, and unusually formatted texts may be harder to detect. Privacy Guardrail is therefore designed as an assistive protection layer—not as a guarantee for complete anonymization, the seamless prevention of data disclosure, or regulatory compliance. This transparency is a key part of the concept: trustworthy AI is built through transparent processes, open documentation, and genuine user control.

Open source with a Clear Beta Boundary

Privacy Guardrail is released as an open-source project on GitHub under the Apache-2.0 license. The open-source code, documented detection rules, and local processing make the system auditable and verifiable. For the first public beta, the extension officially supports Chrome on desktop as well as the platforms chatgpt.com, chat.openai.com, claude.ai, and gemini.google.com. Other Chromium-based browsers may generally work but are not fully tested at this time. Looking ahead, the team is working on better detection quality, smaller and more efficient local models, additional platforms, and potential mobile scenarios.

Secure Authentication Strategies

Passwordless MFA for Modern Cybersecurity