Security and self-hosting
Built to pass your security review.
Signoria runs inside your infrastructure, reads your AI platforms without changing them and keeps a record your auditors can verify. This page sets out how, in the detail your security team will ask for.
Designed for the scrutiny of supervisors, auditors and your own security team
How it is built
Each point describes the platform as shipped. Where a safeguard can be switched off, switching it off is recorded with a reason and shown, never hidden.
Deployment and supply chain
- Installed in your data centre or your cloud account with Docker Compose. Your data stays in your infrastructure.
- Each release is one file, signed by us, that you verify with our public key before anything of it runs.
- The release holds every image the installation runs, so nothing is downloaded while it is installed.
- Each image comes with a software bill of materials, and packages and images are checked for known vulnerabilities before every release.
Identity and access
- Single sign-on with Microsoft Entra ID, under your own multi-factor and access policies.
- People and roles come from your directory groups, synchronised on a schedule. Nobody is created at sign-in.
- Roles decide what each person may see and do, down to the organisational unit. Auditors get read-only access.
- Passwords are for administrators only, with strong hashing, a second factor, a lockout after failed attempts and a limit per address.
Read-only connectors
- Each connector has its own credential, separate from sign-in. A credential that can do more than read is refused.
- Connector secrets are sealed under a key that only your operator holds.
- What a platform reports, such as names, descriptions and instructions, is treated as untrusted text: shown as text, never run as markup or followed as a link.
Audit ledger and evidence
- Every change is entered in a hash-chained ledger in the same database transaction as the change itself.
- The ledger is sealed regularly outside the database and can be verified at any time. A ledger that does not verify is not sealed, and the administrators and auditors are told.
- The platform cannot rewrite its own ledger, and every administrator action is recorded.
- Approvals and assessments never change. A restore from a backup is itself put on record.
Personal data
- Logs hold no request content, names or e-mail addresses.
- Every answer that shows people's names is recorded, so you can see who looked up whom.
- A former employee can be erased on request. What they did stays on record under an anonymous identifier, and the erasure outlasts a restore from backup.
- Uploaded files are scanned for malware before they are kept. Links people add are never fetched by the platform.
AI under your control
- Every AI feature is off as shipped, and each has its own switch.
- The AI features only read. They cannot change a record, and they never approve or classify.
- Where a fact is missing, the platform says it cannot classify rather than guess a risk tier.
- Content goes only to the AI provider your administrator sets up, and its key is sealed.
Safeguards that stay visible
- The approver is never the requester, and changes to workflows and forms are published by a second person.
- These safeguards, the read-only credential and the second factor can be switched off only with a reason. The change is recorded and shown for as long as it stays off.
What leaves the platform
- As shipped, nothing.
- Mail goes through your own mail relay and carries only the asset's code and name, what is asked and a link.
- Notifications in Microsoft Teams, if you switch them on, go through your own Azure bot.
- The AI features, if you switch them on, send only the content they need to the provider you chose.
How we test and fix
- We review the platform against the OWASP Application Security Verification Standard at level 2 and fix what we find, the most severe first.
- Every change to our code is checked for leaked secrets, and everything the build runs is pinned to an exact version.
- We walk your security team through the architecture before anything is installed.
What it does not do
- It does not sit between agents and their users. It does not intercept agent traffic and is not a proxy.
- It does not replace your GRC tool. It is the system of record for AI assets and their approvals, and your other tools read it over the REST API.
See it on your own AI estate.
Book a demo on example data, or ask about the free 90-day pilot.