Security
Minimisation by default.
Proof without content.
This page is written for your DPO and information-security team: exactly what is stored, what is never stored, how the keys are handled, and how any third party can verify the evidence chain without trusting us.
What every log entry contains
- Timestamp and request ID
- Which enforcement point the call came from
- Classification: which types of personal data were found (e.g. national ID, name), never the values themselves
- The decision: masked, blocked or passed on
- Salted SHA-256 hashes of the original and the cleaned content
- Chain fields: previous entry's hash, own hash and an Ed25519 signature
What is never stored
- The prompt content itself
- Patient or citizen data in plaintext
- The detected values: national IDs, names, addresses
- The AI model's responses
- Keys, passwords or tokens from the traffic
We log what was decided, not what was seen.
All content hashes are salted, so a national ID cannot be brute-forced from its hash. A log entry can prove that specific content passed through, without anyone being able to derive the content from the entry.
The chain and the keys
Three mechanisms, one guarantee.
The hash chain
Every log entry is hashed (SHA-256) together with the previous entry's hash. Alter or delete one row, and the chain breaks visibly for every subsequent entry.
The signature
Every entry is signed with Ed25519 across the entry's hash and the previous hash. The signing keys are generated and kept by you, on your own network.
The time anchor
Optionally, the chain is continuously anchored with an independent RFC 3161 timestamping authority, so not even the timestamps can be rewritten in hindsight.
Independent verification
Don't trust us. Verify us.
The evidence layer stands outside the systems it documents. Every export is accompanied by a chain-integrity certificate, and any third party, a regulator, an auditor or a counterparty, can verify the chain and signatures with the public key and standard tooling alone.
Deployment architecture
Built on zero-trust principles.
100% on-premise, no call-home
The entire stack runs on your infrastructure. There is no telemetry, no cloud dependency and no connection back to us in operation.
Closed internal communication
The services talk over Unix domain sockets with locked file permissions, not over the network. Only a single endpoint is exposed externally.
No secrets in environment variables
Keys and secrets are delivered as mounted files with strict permissions, never as environment variables readable by other processes and tools.
Least privilege
The containers run without privileged rights and with shared, locked group access to the internal sockets, following the principle of least access.
Designed to support GDPR art. 5(1)(c) and art. 32-33, EU AI Act art. 12, NIS2 and ISO 27001 controls. Compliance is achieved in your overall setup; we provide the technical documentation and the evidence chain.
Ready to take control of your AI infrastructure?
CareProxy is under active development. Join the waitlist to be notified as soon as pilot installations open.
Or reach out directly kontakt@careproxy.dk