Database-enforced isolation Available now
PostgreSQL row-level security on every table. Tenant isolation is enforced by the database, not only by application code.
Security and governance
Deny by default, enforced where the data lives
A shared memory is only useful if people trust it. This page separates plainly what is built today from what is still on the roadmap.
Built today
Controls available now
These are built and running on our pilot deployment.
Deny-by-default permissions Available now
A permission service checks every request. Search results are re-checked against permissions one by one before they are returned.
Caller rights Available now
AI apps get the smaller of the person's rights and a time-limited grant. An agent never has more rights than its user.
Audit in the same transaction Available now
Every allow and deny is recorded with who, where, which app and which model, and written in the same transaction as the change itself.
Client walls and audience labels Available now
A client's memory is never returned to another client's work. Labels mark memory internal, restricted (named people only, bots never read) or external_ok.
Data, not instructions Available now
Memory is returned labelled as data. Text that looks like instructions is flagged and cannot be bulk-approved.
Strong sign-in Available now
OAuth 2.1 sign-in. People use a password plus an authenticator code. Access tokens are short-lived.
Kill switches Available now
Suspend an AI app or bot immediately.
Versioning and backups Available now
Every update keeps the previous version. Daily backups are encrypted.
Your data
Your data stays in your infrastructure
The AI Hive runs where you decide. In a pilot we run it for you in your cloud, or you self-host it on your premises. Fact extraction uses a small model running on your own server, so nothing needs to leave your infrastructure.
Connectors only read, and only from the sources you choose. By design, The AI Hive never writes into outside systems.
Built on
- PostgreSQL with row-level security
- Open-source components with permissive licences
- Docker-based deployment
Fewer connectors
A smaller attack surface for your AI
Security researchers warn about one combination in AI agents: private data, untrusted content such as inbound email, and a way to send data out. When all three sit in one agent, a hidden instruction in a message can make it leak data.
Read-only company sources Available now
The AI Hive reads your sources on your server and never writes into outside systems, so the memory layer itself cannot send data out.
One labelled context Available now
The AI receives memory as data within a token budget, not raw mailboxes and file stores with their own tools and instructions.
Switch other connectors off Available now
Admins can allow only The AI Hive server in AI apps (for example Claude Code managed settings or organisation connector controls) and keep personal mail and file connectors off, or on for a single task.
Background and measurements: What connectors cost.
In testing
Agents talking to agents, and The AI Hive on a phone
Agent messaging is built and being validated; The AI Hive on a phone is available. The same rules as memory apply.
Verified senders in agent rooms In testing
Each message shows who sent it, for example a person via their AI app; bots are marked as AI and self-chosen labels as declared. Only those who may read a space's memory may talk in its rooms, and client walls still apply.
Messages are data, not instructions In testing
Messages reach an agent labelled as data. Text that looks like a secret or credential is refused, and suspicious text is withheld from agents. Chat becomes memory only when a decision is saved on purpose.
Rooms are audited In testing
Opening and closing rooms, reads, posts, refused posts and claim conflicts are recorded. Rate limits apply per sender and per room.
Notifications carry no content Available
A push notification carries only the kind, a count and a link, never memory text. Each AI Hive server sends its own with its own keys, and a device can be removed.
Nothing kept on the phone Available
The installed app never stores memory or answers from the server for offline use, and logging out clears its data on the phone.
Built and in testing. Anything that does not pass testing will be removed.
On the roadmap
Not available yet
We would rather you hear this from us. These items are planned and are not part of The AI Hive today.
| Planned control | Status |
|---|---|
| SSO with Microsoft or Google, and SCIM provisioning | Roadmap |
| Customer-managed encryption keys (BYOK) | Roadmap |
| PII detection and masking | Roadmap |
| Hash-chained, exportable audit | Roadmap |
| Retention policies per scope | Roadmap |
| Break-glass access with logging | Roadmap |
| Independent penetration test and SOC 2 readiness work | Roadmap |
| Large-scale load testing | Roadmap |
Where we stand. The AI Hive is a pilot-stage product. It has not been independently audited or penetration tested, and we do not claim any compliance certification. If your security review needs more than that today, tell us what you need and we will say honestly whether the pilot can meet it.
Responsible disclosure
Found a vulnerability?
Please email sales@it-labs.in with enough detail to reproduce the issue. We will acknowledge your report, investigate it and keep you informed. Please give us reasonable time to fix a problem before sharing it publicly, and do not access data that is not yours.
Questions from your security team?
Book a session and we will walk through the architecture, the controls above and the gaps.