Amazon Bedrock Guardrails has unveiled its image content filters, introducing top-tier protections that enable users to block up to 88% of harmful multimodal content. This feature is now generally available.
Significant Upgrade in Content Moderation
In a significant upgrade, Amazon Bedrock Guardrails allows for the moderation of both image and text content within generative AI applications. Previously, the service was limited to filtering text only, but this enhancement streamlines the process by alleviating the burdens of creating custom image safeguards and the often tedious manual content moderation.
Milo Carter, VP of Innovative Solutions at KONE, expressed enthusiasm about the potential of Amazon Bedrock Guardrails, stating, “KONE sees Amazon Bedrock Guardrails as essential for safeguarding generative AI applications, particularly in ensuring relevance and context. The integration of product design diagrams and manuals into our applications, supported by Bedrock Guardrails, will enhance accuracy in diagnosing and analyzing multimodal content.”
Customizable Protections for Generative AI Applications
Amazon Bedrock Guardrails provides customizable protections to help users filter out harmful or inappropriate content in their generative AI applications. Users can tailor Guardrails to their specific needs by implementing various policies to detect and eliminate unwanted inputs and outputs. Additionally, the system can identify model hallucinations, ensuring that responses remain grounded and accurate.
The ApplyGuardrail API allows for consistent policy application across any foundation model, whether hosted on Amazon Bedrock, self-hosted, or sourced from third parties. This feature integrates seamlessly with Bedrock Agents and Knowledge Bases, facilitating the enforcement of safeguards across diverse workflows, including Retrieval Augmented Generation (RAG) systems and agentic applications.
Specific Policies Included in Guardrails
The Guardrails platform includes six specific policies, such as:
- Content filters for detecting harmful material across various categories (including hate speech, insults, sexual content, and violence)
- Topic filters to restrict certain subjects
- Sensitive information filters to block personally identifiable information (PII)
- Word filters
- Contextual grounding checks to catch hallucinations
- Automated Reasoning checks (currently in gated preview) to validate, correct, and explain factual claims
With the addition of image content moderation, these safeguards now span both text and images, helping users block significant volumes of harmful multimodal content. Users can configure moderation settings for text, images, or both, adjusting thresholds from low to high to align with their organization’s responsible AI policies.
Availability and Getting Started
This feature is now available in the US East (N. Virginia), US West (Oregon), Europe (Frankfurt), and Asia Pacific (Tokyo) AWS Regions.
To get started with the image content filters in Amazon Bedrock Guardrails, users can create a guardrail on the AWS Management Console and set up content filters for either text, image data, or both.
Steps to Create a Guardrail
- In the Amazon Bedrock console, navigate to Safeguards and select Guardrails.
- Click on Create guardrail.
- In the Configure content filters section, utilize the existing filters to detect and block image as well as text data.
- After configuring the desired filters, save the guardrail to start filtering harmful or unwanted content for your generative AI applications.
Testing the Guardrail
To test the guardrail, select it in the Amazon Bedrock console and choose Test. You can either test it by invoking a model or through the ApplyGuardrail API without invoking a model. The API allows validation of content at any stage in the application flow, ensuring that inputs and outputs are properly evaluated before being presented to the user.
For example, when using the ApplyGuardrail API, you could assess a Meta Llama 3.2 model hosted on Amazon SageMaker or a Mistral NeMo model running locally.
To test a guardrail with image generation, you can generate an image using the Stability model on Amazon Bedrock while incorporating the guardrail.
Further Reading and Resources
This blog offers valuable insights, and for additional reading, check out this related blog post that discusses similar topics. Moreover, for authoritative insights on this subject, refer to this excellent resource.
For further exploration of career skills and training related to Amazon’s initiatives, Fast Company provides a great overview of how Amazon is adapting to evolving workforce needs.
Leave a Reply