Security & data protection

Handing an AI agent access to your customers, your calendar and your ad budget is a real decision. Gridwork is built for that: independently certified infrastructure underneath, workspace isolation enforced by the database itself, encrypted credentials, and an approval gate with an append-only audit log in front of every consequential action. This page describes each control in enough detail to be checked.

Certified infrastructure

Gridwork runs on infrastructure that is independently audited to SOC 2 Type II. These are our providers' certifications, covering the platforms Gridwork is built on:

• Supabase — managed Postgres, authentication and file storage. SOC 2 Type II and ISO 27001; see supabase.com/security.
• Render — application hosting and compute. SOC 2 Type II and ISO 27001; see render.com/security.

On top of that foundation, the controls below are ours, implemented in the product rather than described in a policy document.

Tenant isolation

Every record that belongs to a workspace carries a workspace identifier, and access is enforced by row-level security in Postgres — the database itself refuses cross-workspace reads. Server-side code that runs with elevated privileges goes through a scoped client that binds the workspace on every query rather than issuing raw cross-tenant SQL. Isolation is a database property here, not a filter someone might forget to write.

Encryption

• Traffic to and from Gridwork is served over TLS.
• The credentials for accounts you connect — WhatsApp tokens, ad-account tokens, CRM keys — are encrypted with AES-256-GCM at the application layer before they are written to the database, so a database read alone does not yield a usable token.
• Data at rest is stored on encrypted managed Postgres infrastructure.

The control plane

This is the control that matters most for an agent product. Before a consequential action runs — sending a message, spending on ads, issuing an invoice, writing to a connected system — it passes through a gate that:

• classifies the action's risk;
• evaluates it against your workspace's policy, including autonomy level and spend caps;
• writes an entry to an append-only audit log, enforced by a database trigger; and
• queues the action for your approval when policy requires it.

Spend caps are enforced in code, not documented as guidance: an agent can propose work above your cap, but the action will not execute. Work that costs real money is quoted before it runs.

Access control

Workspace access is role-based, and Gridwork staff access to production is limited to what is needed to operate and support the service. Administrative surfaces are authenticated and restricted to an explicit list of platform administrators.

AI-specific handling

• Your conversations are not used to train shared or public models. Our model provider processes prompts to return a response and does not train on them.
• Knowledge retrieval is scoped to your workspace, and that scoping is applied both in the query and at the database boundary.
• Agents are built not to invent prices, stock or policy. Where they cannot answer from your real data, they are instructed to say so and escalate rather than guess.
• Any code an agent executes runs in an isolated sandbox that fails closed if it cannot be isolated.

Consent

Consent is a first-class field on every contact, not a note in a spreadsheet. Proactive messaging is gated on it in code: a cadence will not send a WhatsApp message to a contact who has not opted in, an opt-out stops the cadence outright, and an unsubscribed contact receives nothing proactive on any channel.

Sub-processors

The third-party services that process data on our behalf — what each one does, and whether it applies to every workspace or only when you connect it — are listed on our sub-processors page. We update that page when the list changes.

Reporting a vulnerability

If you believe you have found a security issue, email security@bygridwork.com with enough detail to reproduce it. Please give us a reasonable opportunity to fix it before disclosing publicly. We will not pursue action against researchers who act in good faith, avoid privacy violations and do not degrade the service.

Vendor reviews

Security questionnaires, data-processing agreements and architecture questions go to security@bygridwork.com.