Security

Last updated: 01 August 2026

How we protect this site, the assistant on it, and the systems we build. Written to be checkable rather than reassuring. Everything below is something you can ask us to demonstrate.

What we are not

We are not certified against ISO 27001, SOC 2, HIPAA, PCI-DSS or FedRAMP. We meet many of the substantive controls those frameworks describe, but we have not been audited against them and we will not imply otherwise. If your procurement requires a certificate today, we are not the right supplier today.

Where data lives

• The website, its database and its backups run on a dedicated server in Germany.
• The language model that answers you runs on hardware we own and operate in Sofia.
• Backups run daily, are encrypted before they are written, and are kept for 30 days.

Access control

• Administrative access to production runs over an encrypted overlay network. There is no public administrative endpoint.
• Server access is key-based only. Password authentication and direct root login are disabled, and repeated failed attempts are banned automatically.
• The firewall denies everything except the traffic needed to serve the website.
• Database roles are per-application, and the databases accept local connections only.

The AI layer

This is where most of the real risk in an AI product sits, and it is where most of our work goes.

• No public model endpoint. The model is not reachable from the internet. Requests reach it only through our application, over a private network.
• Tools enforce permissions, prompts do not. Where an agent can take an action, the permission check lives in the tool, not in the instructions. A cleverly worded message cannot talk its way past it.
• Prompt-injection defence. Input is screened before it reaches the model and output is filtered before it reaches you. Refused attempts are logged and repeat sources are blocked automatically.
• Restricted egress. The machine running the model can reach a named list of hosts and nothing else.
• The assistant cannot rewrite itself. Its instructions and its knowledge base are read-only to it.

Logging and retention

• Server logs: 30 days.
• Assistant conversations: kept only if you consent, then 24 months from last activity.
• Abuse telemetry: a salted hash of your IP address for 24 hours, never the address itself. Refused prompt-injection attempts for 30 days.

Secrets

Credentials are never committed in plain text. Project secrets are encrypted at rest and decrypted only on machines that hold the private key. Server-side secrets live in a root-owned file readable only by the service that needs them.

Reporting a problem

If you find a security issue, write to office@cloudsource.bg rather than exploiting it. Coordinated disclosure earns our thanks and, where appropriate, public credit. Please give us a reasonable window to fix the issue before publishing it.

For systems we build for you

The controls above describe our own infrastructure. For a system we build for a client, security is specified in that engagement: usually your cloud accounts, your identity provider and your retention rules, with the same principles applied. We write it into the statement of work rather than leaving it implied.