AWS Introduces New Policy Guardrails for Safeguarding GenAI
AWS is unveiling new Amazon Bedrock Guardrails policy enhancements, enabling organizations to standardize new protections across their generative AI (GenAI) applications. Amazon Bedrock Guardrails offers a set of robust policies to enhance security standards for AI models, driving protection and responsible AI policies at scale, more effectively.
According to AWS, Amazon Bedrock Guardrails detects harmful multimodal content with up to 88% accuracy, helping to filter sensitive information and prevent hallucinations. The latest policy enhancements further reinforce Amazon Bedrock Guardrails’ efficacy, strengthening content protection across GenAI systems.
The first of these new guardrails, multimodal toxicity detection for image content, enable GenAI models to more comprehensively evaluate, detect, and filter out undesirable or harmful content from both text and image content. Users can configure content filters across a variety of categories—from hate speech to insults, sexual content, violence, misconduct, and prompt attacks—and pair them with configurable thresholds that determine the extent of the model’s filtering.
Amazon Bedrock Guardrails now also features sensitive information protection capabilities with enhanced personally identifiable information (PII) masking for input prompts. This capability enables the detection of PII information—such as names, addresses, or phone numbers—and allows users to block or mask this information. Block mode completely rejects requests containing sensitive information, while Mask mode redacts sensitive data by replacing it with standardized identifiers.
The new IAM policy-based enforcement helps enterprises persist mandatory guardrails for every model inference call, ensuring that safety policies are consistently enforced. This capability provides centralized control across all AI interactions, helping enterprises address common issues such as content appropriateness, safety concerns, and privacy protection requirements.
AWS is further introducing an update to the way Amazon Bedrock Guardrails applies policies, offering more granular control over policy application. Previously, Amazon Bedrock Guardrails applied policies to inputs and outputs by default; now, users can selectively apply guardrail policies to inputs, outputs, or both. This helps enhance performance by reducing unnecessary processing overhead, thereby improving response times without sacrificing protection, according to AWS.
Finally, AWS’ new monitor or analyze mode optimizes configuration, enabling users to determine guardrail effectiveness without having to directly apply policies to their applications. This empowers teams to experiment with different policies and combinations before deployment, driving greater visibility into how guardrails would perform.
“The new capabilities for Amazon Bedrock Guardrails represent our continued commitment to helping customers implement responsible AI practices effectively at scale,” said Esra Kayabali, senior solutions architect at AWS. “Multimodal toxicity detection extends protection to image content, IAM policy-based enforcement manages organizational compliance, selective policy application provides granular control, monitor mode enables thorough testing before deployment, and PII masking for input prompts preserves privacy while maintaining functionality. Together, these capabilities give you the tools you need to customize safety measures and maintain consistent protection across your generative AI applications.”
To learn more about Amazon Bedrock Guardrails’ latest policies, please visit https://aws.amazon.com/?nc2=h_lg.