Security and privacy
How AgentOS keeps each organisation's data separate, keeps people in control of agents, and supports GDPR and EU AI Act work.
Tenant isolation
Every organisation-owned table has a PostgreSQL row-level-security policy that checks membership and role for the signed-in user, so a query can only ever see the rows of organisations that user belongs to. Background jobs, retrieval, analytics and file processing run under the same scoping.
Agents are controlled software
- Agents only receive the data each step passes to them and the knowledge bases they are explicitly linked to.
- External actions go through a server-side tool gateway that checks the tenant, connection, workflow permission, approval, rate limit and idempotency key.
- An approval is bound to a hash of the exact action. If anything changes, the gateway refuses and fresh approval is needed.
- Spending limits, step limits, timeouts and retry limits are enforced by the platform, not by prompts.
Prompt injection
Customer messages and retrieved documents are wrapped and labelled as untrusted data. Even if a model is manipulated, it cannot send, pay or delete anything, because it has no direct access to integrations. Its output is only a proposal that passes validation, policy checks and human approval.
Credentials
Model API keys and integration secrets are stored encrypted in a database vault and read only by server-side code inside the gateway or model adapter. They are never returned to the browser or included in prompts.
Accounts
Passwords are handled by the managed authentication service and never stored by AgentOS itself. Role-based access (owner, admin, manager, employee, viewer) and an append-only audit log are built in. Multi-factor authentication is on the roadmap.
GDPR
Data-subject access and erasure workflows, retention settings, data minimisation in agent prompts, and an export of execution records. Processor agreements and sub-processor registers are maintained by the operator of each deployment.
EU AI Act
A structured self-assessment records intended purpose, deployment context, possible prohibited or high-risk categories and transparency duties. It never marks a system as compliant on its own. Potentially high-risk uses are flagged for qualified legal review.