Security

Rule

Treat each action, route, query, and external request as a trust boundary. Validate inputs, check the caller's authority for every row, and fail closed when access cannot be established.

Validate input and queries

• Define action inputs with a Zod schema; validate data from forms, URLs, providers, and files before using it.
• Build database queries with Drizzle. For raw SQL, bind values as parameters; never concatenate user input into SQL text.
• Keep large uploads and file contents in file storage. Store only their returned ids or URLs in SQL.

Authorize rows and actions

Choose a table's access model when designing its data:

Data	Access model
Private user rows	Store owner_email and filter every read and write by the current caller.
Organization-wide catalog	Store org_id and scope action queries to the active org; gate restricted mutations by org or app role.
Team-shared or individually shareable resources	Use ownableColumns(), register the resource type, and check access with the sharing helpers.
Child rows	Check access to the parent before reading or changing children; add any child-specific rules in the action.

Do not make a team-wide catalog private by adding a per-user owner filter. For an organization-owned catalog, org_id without owner_email is valid: scoped database tools expose only the active organization's rows, while app actions still need their own org filter and authorization. For resources users can share, use these helpers:

• Add accessFilter(table, sharesTable) to list and read-many queries.
• Use resolveAccess(type, id) for an optional read by id, or assertAccess(type, id, role) before a read or write that requires access.
• Use an action's authorize gate for role-wide restrictions. It decides who may call the operation; it does not replace row-level access checks.

Prefer actions for normal app data. A custom HTTP route is appropriate for protocols such as uploads, streaming, webhooks, OAuth callbacks, or public unauthenticated resources. An authenticated custom route must establish the session/request context and apply the same row-level access checks as an action.

Handle untrusted content and URLs

• React escapes rendered text. Do not pass untrusted content to dangerouslySetInnerHTML, innerHTML, eval, or document.write.
• Use ssrfSafeFetch from @agent-native/core/extensions/url-safety for server-side requests to user- or agent-controlled URLs. It checks private network destinations and redirects.
• Treat provider error bodies and uploaded files as untrusted input. Return a useful, bounded error to the user without exposing internal details.

Protect credentials and consequential actions

Never put credential values in source, prompts, action arguments or results, application state, client bundles, logs, or error messages. Use the shared credential flow in the secrets skill.

For a genuinely high-consequence outward action, such as sending a message, charging a payment method, or publishing publicly, use needsApproval to ask the user to approve that call. Approval does not grant access: keep the normal authorize and row-level checks as well.

Related skills

• actions — define validated operations shared by the agent and UI.
• secrets — connect external services and resolve credentials safely.
• sharing — configure access to shareable resources.