AWS Introduces Guardrails for Amazon Bedrock to Enhance the Security of Large Language Models (LLMs)

In today's tech landscape, discussions surrounding the business benefits of large language models (LLMs) are common. However, these models come with several well-known challenges. Implementing measures to control their output is essential for leveraging these powerful technologies responsibly. During the AWS re:Invent conference in Las Vegas, AWS CEO Adam Selipsky introduced Guardrails for Amazon Bedrock.

"With Guardrails for Amazon Bedrock, you can effectively establish safeguards that ensure relevant and secure user interactions, all while aligning with your company's policies and values," the company stated in a blog post.

This innovative tool empowers businesses to define and restrict the types of language their models can generate. If a user poses a question that isn't pertinent to the bot's purpose, it won’t provide an answer. This approach helps prevent delivering misleading responses, or even worse—offensive content that could damage a brand's reputation.

At its core, the feature enables companies to designate specific topics as off-limits for the model. For example, a financial services firm may want to prohibit the bot from giving investment advice, recognizing the risks associated with inappropriate recommendations. A typical scenario could look like this:

"I designate a restricted topic labeled 'Investment advice' accompanied by a natural language description, such as 'Investment advice includes queries, guidance, or suggestions about managing or allocating funds or assets to achieve financial goals or generate returns.'"

Moreover, users can filter out certain words and phrases to eliminate potentially offensive content, applying varying levels of restrictions to convey which language is unacceptable. Additionally, the ability to filter out personally identifiable information (PII) ensures that sensitive data remains protected within model responses.

Ray Wang, founder and principal analyst at Constellation Research, emphasizes the importance of this tool for developers working with LLMs to manage unwanted outputs. "One of the major challenges is creating responsible AI that is both safe and user-friendly. Content filtering and PII protection are two of the top five issues developers encounter," Wang noted. "Ensuring transparency, explainability, and reversibility is essential as well," he added.

The Guardrails feature is currently in preview and is expected to be available to all customers sometime next year.