Security guardrails
Updated on 3 November 2026.
Scan what agents send to an MCP server and what the server sends back, and block, redact or flag sensitive data and attacks.
Request and response
Actions
Choose the instructions for the mode that governs your MCP Gateway. See Switching from classic settings for what happens to stored settings when the engine is on.
- Classic settings
- Policy Engine
Configure it on the server
Open Settings > AI Gateway > MCP Gateway, then the server's Configure > Guardrails section.
The Actions card has Default, Request (tool arguments), and Response (tool results) selectors. Default starts as Monitor; choosing Default for a direction uses that value.
Categories
Enable the data-risk and adversarial categories you need. Each category can have its own Request and Response action; Default uses the Actions card. The Coverage and action summaries show what is selected.
Click Save settings to apply. For differences by person or Smart Group, use Group and user rules.
Going further with the Policy Engine
Open Govern > Policy Engine > MCP Gateway > Controls > Data & Adversarial Risks. It has two sections:
- Choose a recipe or start a blank rule in the appropriate section.
- Choose Applies to, the data types or detectors, and the outcome.
- Check Reads as, then Save to draft.
- Use Test > Try a situation or Replay activity to check the candidate, then review and publish.
Detection rules resolve a payload-wide action: the highest numeric priority wins, with the stricter action winning ties. For each finding, that action competes by priority with the scan profile's category-specific or fallback action. Do not assume every Block overrides every higher-priority rule.
Scan-profile settings resolve per detector and per category. On wins a detector tie, and the stricter category action wins an action tie; the default action is a fallback. Detectors run only when a rule or profile selects them. Without selected checks, nothing is scanned automatically.
Rules that select detectors or set scan-profile actions run before scanning. Put conditions on Data found in a separate detection rule, rather than making the scan depend on a finding it has not yet produced.
Open Activity for observed matches. Classic settings marked Controlled by Policy Engine are stored for use when the engine is off; Edit anyway does not change published policies.
Where detections show
- Overall analytics > Data & security shows flagged request and response content in MCP Gateway Insights.
- Findings & Interactions lists detections and supports input/output DLP filters.
- The server card's Guardrails chip opens its configuration.
Related
- Group and user rules - more or less restrictive guardrail actions for a smart group or user.
- Tool visibility - turn tools off or require confirmation.
- Detection models - what counts as sensitive data or an adversarial prompt.
- Policy Engine overview - how cards, stages and priorities work.