Protect sensitive data before it reaches the AI model.

The exposure point

The critical moment happens before inference.

Once an original sensitive value has been sent to an external model, a later warning or redaction cannot reverse that disclosure. Protection therefore needs to happen in the request path before the model receives the content.

Sano Work is designed to detect supported sensitive spans locally and replace them while preserving enough surrounding meaning for the AI task to remain useful.

• Protect before transmission
• Preserve useful surrounding context
• Reduce original sensitive values in prompts
• Keep the protection step close to the user

Sano Shield 1

A privacy model measured for German and English.

Sano Shield 1 is the detection model behind the protection layer. It is designed for privacy-relevant language patterns and evaluated on held-back examples in German and English.

The model card is published on Hugging Face so technical teams can inspect the stated evaluation approach, limitations, and release information rather than relying only on marketing claims.

• German and English sensitive-span detection
• Published model card and evaluation context
• Purpose-built for the protection layer
• Designed to run close to the user

Defense in depth

Privacy protection is one layer, not the entire policy.

Local replacement supports data minimization, but organizations still need appropriate access controls, provider terms, legal bases, retention decisions, user guidance, and security review for each use case.

Sano Work makes one important boundary technical and repeatable. It should be used alongside the wider governance and contractual controls appropriate to the organization.

• Combine technical and organizational controls
• Review use cases and model providers
• Limit access to protected workflows
• Document responsibilities and exceptions

Protect, infer, restore.

01 · Protect

Supported sensitive spans are detected and replaced locally before the external request.

02 · Infer

The chosen AI model receives placeholders and the surrounding context needed for the task.

03 · Restore

Original values are restored locally in the response for the authorized user.

AI data protection questions

Why not simply ask employees to remove sensitive data manually?

Policies and training remain important, but manual removal is inconsistent and adds friction to every request. A local technical layer can make supported detection and replacement more repeatable before content reaches a model.

Does Sano Shield 1 detect every possible sensitive value?

No detection model is perfect. The published model card describes the evaluated scope and limitations. High-risk use cases should combine detection with access controls, review, and other appropriate safeguards.

Where can technical teams inspect Sano Shield 1?

The public Sano Shield 1 multilingual model card is available through the SanoAI organization on Hugging Face and is linked from the main Sano Work site.

Use the protection layer in context

Privacy-First AI Workspace

See how local protection fits into everyday AI work.

Secure AI for Enterprise

Plan protected AI adoption across teams and organizational controls.

Reduce sensitive exposure before the next AI request.

Talk to Sano about the data patterns, languages, models, and workflows your organization needs to protect.