Security centre

Security and privacy: we don't ask for trust, we give you the means to verify.

AnonyAI is built around a single promise: no document leaves the workstation. This page explains how the architecture guarantees it, rather than a policy, and how your IT department or DPO can check it for themselves.

Local processing

Extraction, detection, review, anonymization and export all run inside the AnonyAI process on the workstation. No step opens a connection.

No telemetry

No usage statistics, crash reports, analytics or “anonymous usage data”. It is not a setting to switch off: it simply does not exist.

Modern cryptography

Restore maps encrypted with ChaCha20-Poly1305, keys derived with scrypt. Licences signed with Ed25519.

Air-gap ready

Works on workstations that are never connected to the Internet, licence activation included.

Data flows

Everything that leaves the workstation. The complete list.

The only outbound traffic concerns the licence, and it happens only when you request it, activate it or check for updates. On an isolated workstation there is no traffic at all.

DataWhenWhyRetention
Email addressTrial, activationIssuing the licence (contract)Duration of the relationship
Company nameTrialIssuing the licence (contract)Duration of the relationship
Licence keyActivation, releaseVerifying the purchaseDuration of the licence
PC fingerprint (salted hash)Trial, activationBinding the licence to a workstationDuration of the activation
Computer name, operating systemActivationRecognising your own PC in supportDuration of the activation
IP address and countryTrial, activationFraud prevention90 days, then deleted
Version numberUpdate checkOffering a new versionNot stored

Never transmitted, under any circumstances

There is no code path, setting or “support mode” that sends:

  • a document or any part of one
  • a file name or path
  • the extracted text
  • a detected value, a placeholder or a map
  • the fingerprint (hash) of a document
  • the number of values found in a document

Why the last two as well

They are the kind of thing a vendor would consider harmless. They are not.

A document fingerprint is an identifier linking a specific file to a specific customer. A count of findings reveals the shape of the document: three people, one IBAN, two tax codes.

Both stay on the workstation, in the local audit log.

Verifiable

Only one module talks to the network. You can check it.

The application has a single point from which a connection can start: the licence client. It exposes five operations, and none of them accepts document content as a parameter.

A reviewer can confirm that there is exactly one network call in the whole code base, and observe with any monitoring tool that no connection is made while a document is being processed.

The application includes no third-party libraries: no dependency that could introduce network behaviour without your knowledge.

# Every network connection in the application $ grep -rn "urlopen\|socket.socket\|create_connection" anonyai/ anonyai/licensing/client.py: urllib.request.urlopen( # One. In the licence client. # Integrity of the downloaded package > certutil -hashfile AnonyAI-win64.zip SHA256 SHA256 hash of AnonyAI-win64.zip: … compare it with the one published on /download
Cryptography

Conservative choices, declared parameters.

No proprietary algorithms. Public standards, documented parameters.

Restore map

Authenticated encryption with ChaCha20-Poly1305 (RFC 8439). Key derived from the passphrase with scrypt (N=215, r=8, p=1): about 100 ms and 32 MB per unlock, to make brute-force attempts expensive.

The cleartext header is bound to the content as associated data: the encrypted content of one map cannot be transplanted onto another.

Licences

Certificates signed with Ed25519 and verified locally against a public key built into the application. Verification needs no network, today or ten years from now.

The private signing key is never present on workstations.

PC identification

Hardware identifiers are read locally and immediately transformed with SHA-256 and a product-specific salt. The raw values never leave the machine and are never stored.

The salt prevents the fingerprint from being correlated with one computed by other software.

No passphrase recovery. If you lose the passphrase of a map, the map is lost: there is no back door, not even for us. The application tells you so before you choose it, not afterwards.
Compliance

AnonyAI's position under the GDPR.

Stated precisely, because a DPO who is told otherwise will find out at the worst possible moment.

Pseudonymisation, while the map exists

A document processed by AnonyAI is pseudonymised within the meaning of Art. 4(5) GDPR as long as you keep the restore map: the data can be re-linked using information kept separately and protected.

Anonymization, once the map is destroyed

Without the map the document becomes anonymous and, as Recital 26 indicates, falls outside the scope of the Regulation. The choice is yours and it is explicit: saving the map is a separate action, with its own passphrase.

Minimisation and privacy by design

Removing personal data before using an AI service is a concrete measure of data minimisation (Art. 5) and data protection by design (Art. 25), and it reduces the personal data transferred to third parties.

Accountability

The local audit log documents the operations carried out on the workstation, supporting accountability (Art. 5(2)), without ever containing the data processed.

And the AI Act? Regulation (EU) 2024/1689 adds to the GDPR; it does not replace it. For organisations using AI it requires, among other things, adequate staff skills (Art. 4) and, for high-risk uses, control over input data. Reducing the personal data that enters AI systems makes both obligations simpler. Read the full guide to the GDPR and the AI Act →

AnonyAI is a tool that supports compliance: the assessment of the processing as a whole remains with the controller. None of this information constitutes legal advice.

Documentation

Material for the DPO and for IT.

We provide the technical documentation needed for vendor assessment, the record of processing activities and any data protection impact assessment.

  • Description of data flows and legal bases
  • Architecture of the application and of the licensing service
  • Cryptographic details and key management
  • Activation procedure for isolated workstations

Vulnerability reporting

If you believe you have found a security problem in AnonyAI or in the licensing service, write to us before making it public. We respond to every report and credit those who disclose responsibly.

Contact
Scope
Desktop application, anonyai.app, api.anonyai.app
Please include
Version, steps to reproduce, impact

Put it to the test in your own environment.

Download it, verify the SHA-256 fingerprint, unplug the network and try it. It is the most convincing test we can offer you.