Security

Built so that one organisation can never see another's risk record.

This page lists the controls that are implemented today and those that are planned. AgentCover holds no security certifications and has not yet been independently audited.

Implemented today

Controls in the current release.

Organisation-level isolation

Row Level Security is enabled on every table. Queries return only rows from organisations where the signed-in user has an active membership.

Server-side authorisation

Role checks run on the server and are enforced again inside the database. A role sent from the browser is never treated as proof of access.

Controlled privilege changes

Memberships and roles can only be changed through audited database functions. Users cannot edit their own role, and platform administrators can only be assigned directly in the database.

Authentication

Email and password sign-in with email verification and password reset, using Supabase Auth with secure, HTTP-only session cookies.

Invitation safety

Invitation tokens are stored only as SHA-256 hashes, expire after seven days and can be accepted only by the invited, verified email address.

Audit log

Organisation creation and updates, invitations, role changes and removals are written to an append-only log visible to owners and admins.

Input validation

Every form is validated on the server with the same schema used in the browser, and constrained again by the database.

Secret handling

No secret keys are shipped to the browser. The application runs entirely with the signed-in user's own permissions.

Planned

Not yet implemented.

Multi-factor authentication and single sign-on.
Shared-store rate limiting across all server instances.
Data export and deletion tooling for organisations.
Encrypted document storage with per-organisation access policies for policy uploads.
Independent security testing before any production launch.

Responsible AI

How AI features will be governed.

No AI model is connected to the platform yet. These are the commitments the AI modules are being designed around.

Human approval for consequential actions: AI analysts will propose; people will decide. Workflows will include explicit approval checkpoints.
Reasoning and sources: Assessments and answers will show how they were reached and which documents they rely on.
Traceability: AI executions will be traced, with cost and latency recorded, and outputs evaluated automatically against test sets.
No overstated authority: AI output will be presented as decision support. It will not be described as a coverage decision, legal advice or an actuarial result.