SafeRedact
For individuals Set up your organisationSet up
Security architecture

Original files never leave your browser

Only the text extracted from each document reaches our API, together with the details of the data subject you enter for the case. Files are never uploaded, and nothing sent for detection is used to train AI models.

Files stay in your browser

Documents are opened, parsed, redacted and exported in the browser. Original files are never uploaded to any server.

No AI training

On the standard route, Anthropic never uses the text for model training and deletes it within 30 days. On the UK and EU route, AWS does not store or log it and does not use it for training.

No document storage

Document content and extracted text are never written to a SafeRedact database or disk. The records we do keep contain no document content and are listed below.

Data flow

Where your data goes

01
Browser

Files opened and parsed

Mailbox exports (PST, MSG, EML), PDF, Word and Excel files are opened in the browser. Nothing leaves the browser at this stage.

02
Browser

Text extraction and pattern matching

Text is extracted locally, and pattern matching in the browser finds structured identifiers such as email addresses, phone numbers, postcodes and National Insurance numbers.

03
In transit

Detection request

The extracted text is sent over TLS to our detection API in London, with the data subject details entered for the case (such as name, email address, phone number, reference numbers and other names) so that the subject’s own information is not redacted from the disclosure. The API holds them in memory only and does not store or log them.

04
In transit

AI classification on your organisation’s route

Standard route: Anthropic’s Claude API classifies the text. Anthropic stores API data in the United States and may process requests there or in selected countries in Europe, Asia and Australia. Under Anthropic’s commercial terms, it is not used for model training, and inputs and outputs are deleted within 30 days.

UK and EU route: Claude on Amazon Bedrock classifies the text in European regions, including London. AWS does not store or log it, does not use it for training, and does not make it available to the model provider.

The route is set per organisation and fixed for each case when the case is created, and every routing change is recorded in an audit trail.

05
Browser

Results returned for review

The detected items return to the browser and are kept locally in IndexedDB while your reviewers work through them.

06
Complete

Redacted pack downloaded

The redacted files, audit trail and review log are produced in the browser and downloaded to your device. SafeRedact holds no copy of them.

Data boundary

What stays in your browser and what is sent

Stays in your browser
Original document files and attachments
Rendered page images
Redacted output files
Detection and review progress (IndexedDB)
Audit trail and review log, until you download them
Data subject records you save for reuse (IndexedDB)
Sent for detection
Text extracted from each document, never the file itself
The data subject details entered for the case
Held in memory by SafeRedact, never stored
Never used to train AI models
Standard route: deleted by Anthropic within 30 days
UK and EU route: not stored or logged by AWS
Encrypted in transit with TLS 1.2 or higher

Records

What SafeRedact keeps

None of these records contains document content, extracted text or file names. They are the records set out in Section 4.2 of the data processing agreements.

Accounts
Names and email addresses of users, organisation membership, agreements and their acceptance, and billing and payment records.
Each file processed
Job and file identifiers, the user’s account identifier, character and byte counts, and a one-way SHA-256 fingerprint of the extracted text, used at export to confirm that the reviewed text is the text that was measured. Each case also has a reference, a keyed one-way hash made in your browser that cannot be reversed to the case name, and a record of its processing route.
Usage and security
One usage entry per user per day of use, and a security log of sign-ins, access and agreement acceptances that includes IP address and browser user agent. Changes of processing route are recorded too.
Help assistant
Where it is enabled for your organisation, a record that a question was asked. The question text is not kept.
Staff notifications
Notices to SafeRedact staff of account events, such as sign-ins, set-up, uploads, completed processing and download attempts, carrying the user’s email address and file counts, and any feedback submitted in the application. These are sent by Slack message and by email.
Emails to users
Copies of the emails we send to users, such as sign-in links, agreement copies, receipts and case notices, and of the staff notices sent by email, kept by our email provider, Resend.

The records held in our database are deleted within 30 days of termination or of a written request to support@saferedact.app, except billing and payment records, which are kept for as long as the law requires.

Sub-processors

Sub-processors

Organisations receive at least 30 days’ notice before a sub-processor is added or replaced. The certifications shown belong to the sub-processors, not to SafeRedact. The full sub-processor schedule sets out the data each one handles and where. Last updated 22 October 2026.

Anthropic, PBCAI classification, standard route; help assistant where enabled · Stored in the United States; processing may also occur in Europe, Asia or Australia · ISO 27001, ISO 42001, SOC 2 Type II
Receives the extracted text and the case’s data subject details for classification, and help questions where the assistant is enabled. Not used for model training; deleted within 30 days.
Amazon Web Services, Inc.AI classification (Amazon Bedrock), UK and EU route · European regions, including London · ISO 27001, SOC 2 Type II
Receives the extracted text and the case’s data subject details for classification. Does not store or log them, does not use them for training, and does not make them available to the model provider.
Vercel Inc.Application and API hosting · AWS Europe (London) · SOC 2 Type II
Serves the application and runs the API that passes text to the AI sub-processor. Request metadata such as IP address; extracted text in transit only.
Supabase Inc.Authentication and database · AWS us-east-1 (United States) · SOC 2 Type II
Holds the account, file, usage, security and route records described above. Row-level security is enforced.
Stripe, Inc.Payment processing · United States · PCI DSS Level 1, SOC 2 Type II
Billing contact and payment records.
Plus Five Five, Inc. (Resend)Email to users · AWS us-east-1 (United States) · SOC 2 Type II
Sends the emails described above and keeps copies under its standard retention.
Slack Technologies, LLCNotifications to SafeRedact staff · United States · SOC 2 Type II, ISO 27001
Carries the staff notifications described above, in SafeRedact’s own workspace. No document content.

Compliance

Regulatory posture

UK GDPR and EU GDPR

SafeRedact acts as processor under Article 28. The UK and EU data processing agreements cover processing on instructions, confidentiality, sub-processor authorisation and notice, breach notification without undue delay, deletion on termination, and audit rights. Download the agreements

CCPA / CPRA

SafeRedact acts as a service provider. Personal information is processed only to detect and redact it, and is never sold, shared for cross-context behavioral advertising, or used for any other purpose.

Sub-processor assurance

Every sub-processor holds a SOC 2 Type II report. These are the sub-processors’ own reports, not SafeRedact’s. Because document content is never stored, there is no document content at rest to secure.

Information security policies

SafeRedact maintains 13 documented information security policies, covering access control, incident response, change management, vendor management, data classification, business continuity, cryptography, logging and monitoring, and privacy. The policy documents are published at saferedact.app/enterprise/policies.

Security hardening

The site is served over HTTPS only, with TLS 1.2 or higher. API endpoints are rate limited. Security headers are deployed on every page: Content-Security-Policy, X-Frame-Options DENY, X-Content-Type-Options nosniff, Strict-Transport-Security with preload, Referrer-Policy strict-origin-when-cross-origin, and Permissions-Policy. The application’s code, libraries and typeface are all served from saferedact.app.

Questions about our architecture?

We're happy to walk through our security model with your privacy or security team.

Contact Sales Enterprise Overview →