Security Gateway

The Security Gateway checks every tool call an employee’s AI client makes against your rules, and logs, redacts, or blocks on a match. Rules for model prompts and completions live in Model guardrails.

How a tool call is scanned

The employee’s client calls a tool.
Entity and regex rules, plus AI Guardrails that screen requests, check the tool arguments.
If not blocked, the Connector sends the call (possibly redacted) to the third party with the employee’s credentials.
AI Guardrails that screen responses check the result before it reaches the client.

Every detection is recorded in Alerts, whatever the action.

Rule types

Rule type	Matches	Use it for
Standard Entity Rules	Pre-trained detectors: names, emails, government and financial IDs, payment data	Regulated data categories
Custom Regex Rules	Your regex, with optional context keywords	Your own identifiers, such as customer IDs
AI Guardrails	A plain-English instruction judged by an evaluator	Topics with no recognizable shape. Can also screen tool results.

Log, redact, or block

Action	What happens
Log	The call goes through unchanged. The detection is recorded.
Redact	The matched value is replaced with a placeholder, such as `[REDACTED:EMAIL]`. The third party never sees the original.
Block	The call is canceled and the client gets an error. The third party is never contacted.

AI Guardrails support only Log and Block.

Override rules for one Tool Pack

Rules apply organization-wide. On a Tool Pack’s Rules tab you can change a rule’s action, add an entity, or switch one off for that pack only. For example, block government IDs everywhere but redact them on the payroll pack.

Violations and alerts

See what your rules catch

Guardrail tester

Test a rule before it sees real traffic