Docs

Security and privacy

Security

What IA4AI stores and never stores, how each layer is enforced, how the service is run today, and what is still to come.

IA4AI records usage, not content. By default it never keeps a prompt, an answer, a file or a tool argument. The layers below are enforced in code and covered by tests, not only written in a policy. The operations section says how the service is run today, checked on the server, and what is not in place yet.

#Security layers

Layer What it protects How it works
Usage only, by default Prompts, answers, files, tool arguments Only declared usage fields are stored. Anything else that arrives (a tampered client, a bug) is dropped on arrival. The extension reads prompt text only to count its characters.
Provider API keys The keys that pay OpenAI, DeepSeek and others Each key is encrypted under a key derived from your own IA4AI workspace key. We store only a hash of that key, so we cannot decrypt your provider keys at rest. A key is opened in memory only while one of your calls passes through.
Other stored secrets Vendor admin credentials Encrypted under a key-encryption key that only the server's root user can read. Rotation needs no downtime. A missing or malformed key stops the server from starting.
Scoped keys Every laptop and agent that holds a key A workspace key reaches only its own workspace, never another one or an operator route. The extension key can only send usage events and fetch its settings. Issuing a new key cuts off the old one at once.
Approval code The actions a copied key must not do Exporting captured content, replacing the key and changing provider keys need a second secret that is never pasted into a tool. Only a salted, slow fingerprint is stored; five wrong tries lock it for 15 minutes; a reset is done by us and recorded.
Access control Each workspace Authentication is fail-closed: the server refuses to start without it outside a local developer run. Roles are scoped per workspace. Every access is logged.
Tamper-evident records The evidence behind every finding An append-only log, hash-chained and HMAC-signed: a change, reordering or deletion after the fact is detected. Each finding says whether it is measured or estimated.
Redaction at ingest Secrets and personal data inside usage records API keys, tokens, private keys, emails, phone numbers, card numbers, SSNs and IP addresses are replaced before anything is stored. Pattern-based: free-form names are not detected.
Content capture, off unless switched on Prompts and answers, when you want a cheaper model verified Only your workspace's admin can turn it on, for 1 to 90 days; it closes by itself. Content is redacted, encrypted with a key unique to your workspace, and deleted after 30 days by default.
Browser extension Employees' browsing No content script: it never reads or changes a page. Its permissions are pinned by a test to the AI chat sites and IA4AI; adding one fails the build. The person can see everything it records in its popup.
Release integrity Production code Code reaches production only after the test suite passes. The server deploys only commits with a verified signature and refuses anything else. A deploy that fails its health check rolls back by itself.
Transport Data in transit HTTPS everywhere: the dashboard, the API, the gateway, and the gateway's calls to providers.

Investigation conclusions are written by a template on our server by default. A language model writes them only if you ask for it, and then only from the findings' sentences (numbers and short labels), never prompts, traces or keys. See Conclusions.

#Where it runs

In the EU, on a server we operate at Hetzner Online GmbH in Falkenstein, Germany. Hetzner is the hosting sub-processor; its Falkenstein data center park is certified to ISO/IEC 27001:2022.

#Operations today

IA4AI is in beta and runs on one server. This is what is in place, checked on the server on 13 November 2026, and what is still to come.

Area In place today Next
Network Two firewalls, at the provider and on the server itself: only SSH, HTTP (which redirects to HTTPS) and HTTPS are reachable; everything else is refused. —
Server access SSH by key only, password login off. A separate admin user, with direct root login switched off.
Encryption in transit TLS 1.2 and 1.3 only, HSTS for one year, certificates renewed automatically. —
Encryption at rest All customer data, the databases and their backups, sits on an encrypted volume (LUKS2). Its key is never stored on the server: after a restart, the volume is unlocked by hand. On top of that, provider keys, vendor credentials and captured content are encrypted by the application. —
Patching Automatic daily security updates. —
Deploys Only signed commits, only after the tests pass; a failed deploy rolls back by itself in 17 to 35 seconds. Branch protection on the repository.
Backups Both databases backed up daily to the encrypted volume and verified (database integrity and the evidence chain); kept 7 days, 4 weeks and 6 months; a weekly restore drill. An encrypted copy off the server, at a Hetzner Storage Box in Germany.
Monitoring Failed deploys, backups and restore drills raise an alert on the server. Alerts sent to a person, and an external uptime check.
Logs Web server logs kept 14 days: IP addresses, never query strings or emails. Application logs capped at 1 GB. —
Dependencies Vulnerability alerts on the code repository. Automated dependency scanning in the test pipeline.
Accounts Private code repository. Two-factor sign-in on every operator account: hosting, code and Google. —
Availability One server, no failover. After a restart it stays down until the encrypted volume is unlocked. Observe-only setups never depend on it; gateway setups can bypass it in about 30 seconds (how). A measured time to rebuild from backup on a new server.
Certifications Our host's data center is certified to ISO/IEC 27001:2022. IA4AI has no SOC 2 or ISO 27001 certification of its own, and no third-party penetration test yet.

#What the evidence log proves, and what it does not

The log is tamper-evident, not tamper-proof. It answers "was this record altered after we stored it?": a changed payload, a reordering or a deletion breaks the chain, and a changed signing key breaks the signatures. It does not defend against an agent that lies at the source (that is why every finding carries its status, and a stated rationale is checked against what the agent did), nor against someone who could re-sign with a stolen key: an independent, external anchor for the log is a follow-up.

#Report a vulnerability

Write to samstern@odux.io, with what you found and how to reproduce it. Security questions for a pilot go to the same address.