Amazon Web Services (AWS) has announced an enhanced capability called Guardrails for Amazon Bedrock, designed to empower developers with customizable safeguards for their AI chatbots and conversational agents. This initiative aims to foster safer, more responsible interactions between end users and generative AI applications.
The updated guardrails integrate seamlessly with Amazon Bedrock, AWS’s platform for building chatbots and voice assistants utilizing large language models. Developers can now set denied topics to restrict specific types of content, configure advanced content filtering across categories such as hate speech and violence, and implement automatic redaction of personally identifiable information (PII) from chatbot responses.
AWS Vice President Antje Barth emphasized the company’s commitment to advancing generative AI “in a responsible, people-centric way.” The guardrails are designed to codify safety and privacy controls tailored to a developer’s unique use cases and responsible AI policies, enabling innovation while effectively managing user experiences.
These guardrails apply to all large language models available through Amazon Bedrock, including fine-tuned models and third-party foundations like Anthropic’s Claude, AI21’s Jurassic, and others. This allows developers to standardize controls consistently across multiple chatbots and assistants.
The denied topics feature enables developers to define undesirable subjects using natural language descriptions. For instance, a banking chatbot could be programmed to avoid providing investment recommendations. The content filters have been enhanced to block outputs at adjustable thresholds for categories such as sexual, violent, hateful, or insulting content.
In addition, Amazon has rolled out automatic PII redaction, which effectively removes or blocks sensitive user inputs like names, emails, or phone numbers, further bolstering user privacy.
The guardrails also integrate with Amazon CloudWatch for comprehensive monitoring. Developers can analyze inputs or responses that violate defined policies, allowing for continuous improvement of their systems.
Currently, the Guardrails for Amazon Bedrock capability is in limited preview. Interested developers can request access through their AWS sales representative or support contacts. The preview supports major AI foundations like Claude, Llama, and Titan Text, with plans to include custom models in the near future.
By providing granular controls over safeguards and policy enforcement, Amazon aims to offer developers greater flexibility while ensuring accountability in responsible AI development using chatbots and conversational AI.
Initial image from aws.amazon.com