Policies

AI Defense runtime protection secures your AI applications with guardrails that block adversarial attacks and harmful responses. Below, we list the runtime guardrails that are available for use in AI Defense policies and Multicloud Defense AI Guardrails profiles.

For each guardrail, we list below the specialized rules that make up that guardrail. For each rule, we explain the rule and list whether it applies to prompt data, response data, or both. Unless noted otherwise, the rules listed here scan both prompts and responses.

At the end of each rule description we provide a threat taxonomy mapping to indicate which industry standards the threat type corresponds to, based on the OWASP Top 10 for LLM Applications 2025 and the MITRE ATLAS Threat Matrix (Adversarial Threat Landscape for Artificial-Intelligence Systems).

The guardrails are as follows:

• Security guardrail

• Privacy guardrail

• Safety guardrail

• Guardrails for Japanese-language content

See Rules and Guardrails to learn about the guardrails.