Amazon Web Services (AWS) has taken a significant step in the field of generative artificial intelligence with the launch of Amazon Bedrock Guardrails, a new service that will be available starting in April 2024. This service aims to address the challenges inherent in using generative artificial intelligence models, offering personalized protection mechanisms to ensure alignment with companies’ responsible artificial intelligence policies.
As generative artificial intelligence models gain popularity for their ability to generate information on a wide variety of topics, they also face critical issues such as content relevance, protection of sensitive information, and prevention of hallucinations, which are errors or misleading information generated by these models. Although Amazon Bedrock already has certain built-in protections, these are often specific to the models and may not fully adapt to the particular needs of each organization.
Developers are often forced to implement additional controls to ensure the security and privacy of their artificial intelligence applications. This challenge intensifies when organizations employ multiple foundational models for various use cases, making it crucial to establish consistent safeguards to facilitate development cycles and promote a unified approach to responsible artificial intelligence.
Amazon Bedrock Guardrails emerges in response to these needs, allowing developers to implement customized safeguards in generative AI applications. These safeguards are not only adaptable to different use cases but can also be applied to multiple foundational models, improving the user experience and standardizing security controls.
One of the most innovative features of Amazon Bedrock Guardrails is the ApplyGuardrail API. This function evaluates both user inputs and model responses, applicable even to custom and third-party models that are not part of Amazon Bedrock. This API is useful in generative artificial intelligence architectures, such as self-hosted language models or Augmented Generation Recovery architectures.
To demonstrate its application, AWS has presented a practical example where a guardrail is implemented to prevent a model from providing financial advice. This example uses filtering policies that address prohibited topics and verify the contextual coherence of responses, ensuring relevance and grounding in valid information for the user’s query.
The announcement of Amazon Bedrock Guardrails represents a notable advancement in the security of generative AI applications, making it easier for companies to incorporate standard and thoroughly evaluated safeguards into their workflows, regardless of the models they use. This reinforces AWS’s commitment to responsible artificial intelligence and optimal protection on its platforms.
via: MiMub in Spanish
