Scout security

Scout is Uniform's AI agent for content authoring. It acts as a human-in-the-loop assistant: it inherits the permissions of the person using it, and every change it makes lands in draft state for review before anything publishes.

What Scout can and can't do#

• Scout has no permissions of its own. It acts as the requesting user and can only see or change what that user can already see or change.
• Scout writes only to draft state. It cannot publish, and it has no access to published content. Workflow stages that auto-publish reject Scout's transitions so a human must perform them.
• Scout can delete entries, compositions, patterns, assets, and content models, but never silently. Pending deletes surface with a distinct treatment and a separate confirm action.
• Deleted entries, compositions, patterns, and assets go to Trash and can be restored for the project's retention period, typically 30 days. Deleting a component definition or content type is permanent.
• We recommend restricting the permissions of the people who use Scout to the entities they need, so a session cannot accidentally delete content types, component definitions, or visitor dimensions.
• Scout cannot manage users, team settings, or billing, and cannot edit, move, or delete existing project map nodes.
• AI features can be disabled at the team level.
• If enabled for your team (developer preview), you can bring your own OpenAI or Azure OpenAI endpoint and API key to control the model and provider directly.

Reviewing Scout's changes before they land#

• Review changes is the default apply mode for new threads. Scout stacks edits as pending edits, and nothing changes in your content until you apply them.
• You can switch a thread to Autonomous if you want Scout to apply edits as they work. The choice is a personal preference, not a team policy.

What happens to your content and prompts#

• Uniform opts out of provider-side prompt storage on model calls. OpenAI retains request logs for 30 days for abuse monitoring only, never for training.
• Your content is never used to train Uniform's or any provider's models. This is a contractual commitment in the DPA.
• Chat threads are encrypted and isolated per user and project, and auto-delete one year after last activity. Users can delete them sooner at any time.
• EU data residency is available for chat storage.

Guarding against prompt injection#

• Scout's instructions and retrieved content are kept in separate channels, so field content is treated as data, not as instructions.
• Content fetched from external URLs is isolated, truncated, and marked as untrusted. Fetching blocks private networks, localhost, and cloud metadata endpoints.
• Scout can only act through a typed, schema-validated set of tools. It cannot take free-form actions. Connected MCP servers and integrations add tools that are equally schema-bound.
• AI-generated output is stored as structured data and passes through the same sanitization as human-authored content before it reaches your site.

Oversight and compliance#

• Scout's edits appear in the same content history as manual edits: who changed what, and when.
• Scout is not classified as high-risk under the EU AI Act. It assists with content, not decisions about people.
• Penetration testing covers the AI agentic features. See the Uniform trust center for current reports.
• Content created or edited through Scout is yours. Uniform claims no IP rights over it.

Model providers and sub-processors#

• Scout runs on OpenAI's current GPT-class models for generation. Contact the Uniform team for the current model list.
• Supporting infrastructure includes Cloudflare (hosting and chat storage), AWS (backend), and PostHog (privacy-mode telemetry, no prompt content).
• Named sub-processors and their processing regions are published in the Uniform trust center and kept current.