Skip to main content
AI Guardrails allow enterprise administrators to define safety boundaries for how users interact with Devin across the organization. Guardrails automatically screen incoming user messages — including initial messages, follow-up messages, and PR comments — to detect prompt injection, data exfiltration attempts, and policy violations before Devin processes them.

​
Overview

Guardrails run as an additional layer of oversight on messages sent to Devin. They analyze user messages in real time and can:
  • Log suspicious messages for review (log_only)
  • Warn the user with a visible banner while still processing the message (warn_user)
  • Block messages that violate organization policies (block_message)

​
Configuring Guardrails

Guardrails are configured at the enterprise level, at Settings > Guardrails, by administrators with the Manage enterprise settings permission. Each guardrail’s setting applies to every organization in the enterprise — guardrails cannot be configured per organization, and organization settings do not have their own Guardrails page. The page has two tabs:
  • Guardrails — The preset guardrails, grouped into categories such as Data controls & privacy and Acceptable Use. Select a guardrail to see its description and example messages.
  • Violations — A chart of violations over time and a table of recorded violations, filterable by guardrail and by organization within the enterprise. Each row links back to the originating session for investigation.
Each guardrail has a single dropdown that both enables it and sets the action taken on violation:
Terminating a session on violation (kill_session) is not a configurable action. The value can still appear on historical violations recorded while it was configurable.

​
Guardrail Events

When a guardrail is triggered, Devin records the event with details including:
  • The user message that triggered the guardrail
  • The guardrail rule that was matched
  • The action taken (log_only, warn_user, or block_message)
  • A link to the session where the event occurred
Guardrail events appear in the audit logs with the ai_guardrail_violation action type, enabling automated monitoring and alerting. You can also retrieve guardrail events programmatically through the guardrail violations API.

​
Use Cases

Common guardrail configurations include:
  • Detecting prompt injection — Identify and block user messages that attempt to override Devin’s instructions or manipulate its behavior
  • Preventing data exfiltration — Flag or block messages that attempt to instruct Devin to send sensitive data to unauthorized destinations
  • Enforcing policy compliance — Screen user requests to ensure they align with organizational security and usage policies
AI Guardrails is an enterprise feature. Contact your account team to learn more about enabling guardrails for your organization.