Security and architecture

Your data stays encrypted, even in our own database.

The database never stores personal data in readable form. Whatever sensitive material exists, exists encrypted under a key kept outside the database: a database administrator, a backup, or an attacker with database access reads nothing. Detection runs on our servers in Germany, before the text leaves masked for the AI.

The data flow

What leaves, what stays, what is never written

What leaves for the AI

The text with the details that were detected replaced by placeholders: [NAME_1], [AFM_2], [IBAN_3], sent to Anthropic (Claude API, USA), the only model provider we use today. The model answers over those. The real values are restored on your screen, inside the response stream. Detection accuracy is measured and published on the accuracy page.

What stays here, encrypted

On our own infrastructure in Germany (Hetzner). The mapping between placeholders and real values, under AES-256-GCM with a key from an environment variable (not in the database, not in the code). Vault search runs on HMAC fingerprints: your query never reaches SQL as plaintext.

What is never stored

Original files (held on temporary disk for up to 30 minutes, 10 minutes for visual redaction, and never in the database), email threads (read live, written nowhere) and the unmasked text of your conversations.

The guarantees

One by one, with the mechanism next to each

Encryption wherever a value exists

An AES-256-GCM value vault, a separate and independent key for HMAC search, and the OAuth tokens of connected mailboxes encrypted too. Keys derive from a master key held in the runtime environment, outside the database and outside the code. In a dedicated installation the master key is yours.

Tamper-evident audit trail

Every audit row gets a sequence number and a keyed MAC over the previous one, forming a chain. Any later edit, delete or insertion breaks the chain from that point and is caught by verification. The key of the chain is not in the database, so not even a database administrator can rewrite history unnoticed.

Fail-closed: when something breaks, we do not leak

If the detection service is unreachable, the request is not forwarded unmasked to the AI. It is refused. Availability is sacrificed before privacy, never the other way round.

Retention on a clock, and on your terms

Original files: on temporary disk, deleted automatically within 30 minutes, never touching the database. The encrypted vault mappings are purged automatically after 30 days. Conversation history (masked text only) stays until you delete it or close the account. For dedicated installations there is a session-only mode, a deployment setting under which no value is written to a database, not even encrypted.

No hidden paths to third parties

Not even the fonts load from external CDNs. Every asset is served by us, so your browser talks to no third party for the app to work. Exceptions you choose: your Google avatar if you sign in with Google, and the Google or Microsoft sign-in pages when you connect a mailbox. What travels to the AI provider (Anthropic, outside the EU) is the masked text, which is the very reason this product exists. Every such send is recorded with destination and time. There is also web search, on by default with a switch in the conversation settings: when the AI decides to search, the query it writes itself runs through the same detection as a message, whatever is detected is removed, and only then does it leave from our server to public search engines; each query and each fetched page is recorded in the audit trail like a send to the model.

Your own key, for individual accounts

On individual accounts that sign in with a password, vault values and content are encrypted under a key your password unlocks. The database holds it wrapped, useless without the password. While the key is locked, not even we can read that data. It unlocks at sign-in and stays in server memory for up to 2 hours of inactivity so the conversation works. It does not apply to organisation workspaces or to Google accounts without a password, which are encrypted under the installation master key.

A human has the last word

A privacy panel on every message (what was hidden, under which placeholder), an approve-before-burn preview on every document redaction, explicit warnings for unreadable handwriting and unsupported scripts. No silent sends.

To be precise

AI processing is performed by a provider outside the EU, over masked text. That is exactly the gap this product closes. The full sub-processor list and processing terms are public, and the measurements behind the detection claims have their own page.

Measured accuracy The team

Responsible vulnerability disclosure

If you find a vulnerability, write to us privately at contact@filterit.ai before publishing and give us a reasonable time to fix it. Do not test against other people’s accounts or data. The contact details are also in /.well-known/security.txt (RFC 9116). We answer every report and credit anyone who wishes on the security page.

Security questions before you start?

For a DPA, vendor questionnaires, or a technical call with your DPO, talk to us. Otherwise, try it free.

Try it free Contact us