JA EN
shiichan

Guard every account at once: Amazon Bedrock Guardrails cross-account safeguards are GA!

AWS Blog ★★★ View Markdown
#aws#ai-safety#ai#infrastructure

Hi everyone, it's me, Shii-chan! Today AWS has a great update for anyone running AI safety controls across a big organization, so let me walk you through it.

AWS Blog aws.amazon.com

What was announced?

On the AWS Blog, AWS announced that cross-account safeguards for Amazon Bedrock Guardrails are now generally available (GA). This lets you centrally enforce and manage AI safety controls (guardrails) across multiple AWS accounts inside your organization.

Until now you had to set up guardrails account by account, but from here you can push one set of rules from a central management account.

The story so far

Bedrock Guardrails filters harmful content and personal information from model inputs and outputs to help you build responsible AI. But in a large organization with many accounts, you had to attach guardrails separately in each one. When you wanted the exact same safety rules to apply everywhere with no gaps, there was no unified way to enforce them.

What changes

This update gives you two levels of enforcement.

  • Organization-level enforcement: one guardrail created in the management account applies to all member accounts and OUs through AWS Organizations policy settings, automatically covering every Bedrock model invocation.
  • Account-level enforcement: guardrails apply automatically to every Bedrock inference call within a specific AWS account.

So even if a developer forgets to attach a guardrail, the safety rules set by the org or an admin still kick in from underneath. Governance folks will love this.

Dive Deep

Here is the rough flow. You start by creating a guardrail with a specific version in the management account.

For account-level enforcement, open the Bedrock Guardrails console, choose Create under Account-level enforcement configurations, and pick which models to include or exclude and how far to guard content (Comprehensive, which enforces on everything, or Selective, which trusts the caller to tag sensitive content).

For organization-level enforcement, enable Bedrock policies in the AWS Organizations console, create an Amazon Bedrock policy that specifies the guardrail ARN and version, and attach it to your target accounts, OUs, or the root.

A few things to keep in mind:

  • Automated Reasoning checks are not supported.
  • If you specify the wrong guardrail ARN, it becomes a policy violation and enforcement will not apply.

It is generally available in all AWS commercial and GovCloud Regions where Bedrock Guardrails is available. You are charged per enforced guardrail according to the safeguards you configure, following Bedrock pricing. If you want the full details, check the cross-account safeguards enforcement developer guide.

Wrap-up

  • You can now centrally enforce and manage Bedrock Guardrails across multiple AWS accounts (GA).
  • Two levels, organization and account, apply your safety rules automatically.
  • You can choose the model scope and how far to guard content (Comprehensive / Selective).
  • Automated Reasoning checks are not supported, and watch out for a wrong ARN.
  • Available in all commercial and GovCloud Regions, billed per safeguard.

If you run lots of accounts and want AI safety enforced across your whole organization with no gaps, this update is a perfect fit for you, governance and security teams!

Related posts