New agent guardrail policies

AI Guardrails protect your agents using a set of predefined categories, but every organization has content, compliance, and data-protection rules that go beyond the defaults. To support your unique business needs, we're evolving Guardrails so you can set custom rules across your topics, terms, and data formats.

Custom guardrail policies

System Administrators can now define three kinds of custom policies that apply as an organizational baseline across every agent:

• Denied topics described in plain language. Detection evaluates meaning and intent, not just keywords, so content is blocked even when it's paraphrased or implied.
• Word and phrase filters for exact terms such as competitor names, internal codewords, or sensitive phrases.
• Custom PII patterns are defined with regular expressions, so sensitive data formats unique to your organization, like employee IDs, account numbers, or booking codes, are detected and either blocked or masked.

Developers can extend this organizational baseline with rules specific to their own agent, adding stricter controls where a use case demands it, but they can never weaken the baseline administrator set. The result is seamless, organization-wide protection with flexibility for targeted, agent-level hardening. Every intervention is logged and traced to the exact rule that triggered it, so it remains easy to audit and troubleshoot guardrail behavior.