Prompt injection guardrails now available in AWS EU, Japan, and APJ regions for Cortex AI

Prompt injection and jailbreak guardrails in Cortex AI are now available when the CORTEX_ENABLED_CROSS_REGION account parameter is configured to AWS_EU, AWS_JP, or AWS_APJ, expanding geographic support for enterprise AI applications.

Use cases

Imagine you manage organizational AI agents and need to ensure malicious users cannot bypass system instructions using manipulated commands. With this feature, you can:

• Detect and block direct and indirect prompt injection attempts at runtime.
• Prevent jailbreak attacks designed to bypass model safety protocols.
• Audit guardrail scan activity and token consumption.

How to use

To enable prompt injection guardrails for your account, use the ALTER ACCOUNT command with the AI_SETTINGS parameter.

ALTER ACCOUNT SET AI_SETTINGS = $$
  guardrails:
    advanced_prompt_injection:
      - enabled: true
$$;

What happens in this code:

1. The command configures AI settings at the account level.
2. The guardrails block enables advanced_prompt_injection protection.
3. The enabled: true flag enforces runtime scanning for CoCo, Snowflake CoWork, and Cortex Agents.

To view your current account configuration:

SHOW PARAMETERS LIKE 'AI_SETTINGS' IN ACCOUNT;

What happens in this code:

1. The command searches existing account parameters.
2. The filter locates the AI_SETTINGS key to verify active guardrails.

To disable the guardrails:

ALTER ACCOUNT UNSET AI_SETTINGS;

What happens in this code:

1. The command removes custom AI_SETTINGS from the account.
2. Cortex AI Guardrails protections are no longer enforced.

Limitations and requirements

• Requires Snowflake Enterprise Edition or higher.
• The account parameter CORTEX_ENABLED_CROSS_REGION must be set to AWS_EU, AWS_JP, AWS_APJ, ANY_REGION, AWS_US, or AWS_GLOBAL.
• Configuration requires the ACCOUNTADMIN role.
• Available for commercial accounts in supported clouds, excluding government environments.
• Usage consumes Snowflake credits based on scanned token volume.