Amazon Bedrock Guardrails: New AI Safety Tiers

Amazon enhances its Bedrock Guardrails with new safeguard tiers, addressing the need for flexible AI safety controls. The update introduces “Classic” and “Standard” tiers for content filters and denied topics, allowing users to tailor protection levels based on application needs. The Classic tier offers low latency but limited language support (English, French, Spanish), while the Standard tier provides multilingual support (60+ languages), improved robustness against prompt attacks, and enhanced topic detection, albeit with slightly higher latency. Cross-Region Inference Support (CRIS) is integrated, enabling better scaling and availability across multiple AWS regions. This tiered approach benefits organizations with diverse AI applications, such as financial services companies deploying AI assistants (Standard tier for comprehensive multilingual protection) and internal analytics tools (Classic tier for low-latency, focused protection). Amazon claims significant improvements in harmful content filtering recall (15%+), balanced accuracy (7%+), and prompt attack protection (30% recall, 16% balanced accuracy) with the Standard tier compared to the Classic tier. The Standard tier also demonstrates superior multilingual performance compared to competitors. Users can configure tiers via the AWS Management Console, SDKs (Python example provided), and APIs. The ApplyGuardrail API allows for efficient testing of guardrails without invoking foundation models, saving costs. Best practices include staged testing, considering language requirements, balancing safety and performance, using policy-level tier selection, and understanding cross-region inference requirements. Overall, the update offers a more nuanced approach to responsible AI, empowering organizations to balance safety with performance and functionality across various applications.

(Source: https://aws.amazon.com/blogs/machine-learning/tailor-responsible-ai-with-new-safeguard-tiers-in-amazon-bedrock-guardrails/)