PostgreSQL security boundary for AI

Bounded PostgreSQL access.
For AI tools, applications, and web APIs.

Apply one set of access rules to AI tools, PostgreSQL connections, and Supabase APIs. Check each request that passes through the boundary, hide configured fields, stop unsafe SQL, keep Supabase's existing row-level permissions where available, and record what each path can establish about the caller.

The problem

AI created new routes to production data.

A careful PostgreSQL role is not enough when an agent can call a different MCP server, a backend holds a broad database credential, or a Supabase client can query protected columns through PostgREST.

Safe Boundary gives each path an enforcement point before data reaches the caller. It complements database grants and RLS; it does not replace them or pretend every caller has verified identity.

Rollout

Observe first. Enforce with evidence.

The boundary only works when the protected caller cannot go around it. Deployment and credential design are part of the security story, not an afterthought.

  1. Map every path

    Inventory database DSNs, Supabase REST URLs, service_role keys, and MCP tools.

  2. Observe real traffic

    Forward unchanged requests while recording would-block and would-mask decisions.

  3. Enforce policy

    Activate blocking and masking after representative clients pass validation.

  4. Close bypasses

    Remove direct credentials and restrict network paths so constrained callers cannot go around the boundary.

Safe defaults today

Zero-trust-aligned enforcement.

Read-only classification, configured masking, dangerous-function denial, inference protection, RLS-safe passthrough, and structured audit are implemented controls. Human SSO can add verified mTLS identity on the wire path.

Do not overclaim

Not yet universal zero trust.

MCP agent labels are attribution rather than verified workload credentials. Policy is not yet signed per workload, quotas are not durable across processes, and direct endpoints must be closed through deployment controls.

Read the security model →

Early access

Start with one real PostgreSQL access path.

Validate client compatibility and policy in observe mode, then enforce with explicit bypass controls. Public packaging is still being validated with design partners.