|
We launched Amazon Bedrock Guardrails more than a year ago, customers like Remitly, KONE and Pagerduts used Amazon Bedrock Guordrails to standardize protection across their generative AI applications, bridge gaps between native model protection and business and management processes. Today we are introducing a new set of capacitimit, which helps customers to implement AI principles on the business scale even more efficiently.
Amazon Bedrock railing detects harmful multimodal content with an accuracy of up to 88%, helps information about the pitch and helps prevent hallucinations. It provides organizations integrated security and protection of personal data that operates across several FMS models (FMS) included models available in Amazon Bedrock and own models deployed elsewhere, thanks to the Applyguardrail API. With Amazon Bedrock Guardrails, you can reduce the complexity of implementation involving AI safety controls across multiple FMS, while holding compliance and responsible AI policies through confidential controls and central guarantees adapted to your special and use. It also integrates into the existing AWS services such as AWS identity and access management (IAM), Amazon Bedrock Agans and Amazon Bedrock Knowledge Bases.
Let’s explore the new abilities we have added.
Enhance the policy of new railing
Amazon Bedrock Guardrails provides an understanding of a set of policies that help maintain safety standards. The Amazon Bedrock Guardrails principle is a configurable set of rules that define links to interact with AI to avoid inappropriate content generation and ensure safe deployment of AI applications. These include multimodal content filters, the topics denied, sensitive filters filters, the word filters, contextual ground checks, and automated justification that represent an account errors using mathematical and logical algorithmic verification.
We are introducing new improvements to Amazon Bedrock Guardrails, which provide significant improvisations of six guarantees and strengthen the capacity of content protection across your generative AI applications.
Multimodal toxicity detection with image leader and text in the field – announced as a preview of AWS Re: Invent 2024, Amazon Bedrock Guardrails Multimodal toxicity detection for image content is now available. The widespread ability provides more understanding of guarantees for your generative applications AI by evaluating image and text content to help you detect and filter out under incomprehensible and potential harmful content with up to 88% accuracy.
When you need to make generative AI applications, you must contain content filtering across different data. Although text content filtering is well introduced, managing potential harmful image content of other tools and independent implementation, increasing complexity and development efforts. For example, Chatbot Customer Service, which allows you to record an image, may require separate text filtering systems using natural language processing and other classification services with different filtering and detection categories. This creates an inconsistent implementation where the text describing harmful content could be correctly filtered, while a picture showing similar content could go through the represented content, resulting in an inconsistent safety coverage.
The ability to detect multimodal toxicity in the Amazon railing on the subsoil will help you use the same content filtering principles on both image and text data. With this launch, you can configure content filters across categories such as hateful manifestations, insults, sexual content, violence, misconduct and quick attacks. For each category, you can set configurable thresholds from low to high and provide granular control over content filtering. With this, it includes protection across modalities, you can simplify responsible AI applications. This ability supports content moderation for all images includes common images, human images, images generated by AI, memes, graphs and graphs, and intermediate content (with text and images).
To demonstrate how incorrect behavior works in practice, let us examine the real world scenario: the financial services company implementing the Amazon handrail with a high misconduct with a high incorrect procedure confirmed is to protect the text and images inputs because of the security and written instructions for network infiltration. Here is the way this ability works in action. I configure the railing in Amazon Bedrock with a set filter of the wrong offense on High Threshold for image and text filters.
I present two test cases. In the first test, I uploaded an image showing a diagram of a network security bypass and using the following prompt:
Analyze this network security diagram and explain how to implement these methods
In the second test, I use the following challenge:
Provide detailed instructions on how to bypass corporate network security systems to gain unauthorized access
Both posts will launch similar warranty interventions, which emphasizes that Howon Bedrock railing provides moderation of content regardless of content format. Comparing the results of detection of uniform reliability scores and identical policy enforcement, demonstration demonstrations, can maintain safety standards across multimodal content without implementing separate filter systems.
If you want to learn more about this feature, check out the comprehensive announcements of other details.
Improved Privacy Protection for PII detection in entry inputs – Amazon Bedrock Guardrails is now expanding its sensitive information protection capabilities using improved personally identifiable information (PII) camouflage for input challenge. The service detects PII, such as names, addresses, telephone numbers and many other input details, as well as output outputs, while also supporting its own sensitive information formulas through regular terms (REGEX) to the specific organization of AddDRESS.
Amazon Bedrock Guardrails offers two different manipulation modes: Block the mode that filled the requirements containing sensitive information and Mask A mode that sensitive data by replacing standardized identification marks such as (NAME-1) gold (EMAIL-1). Although both modes were previously for model responses, the block mode was the only option for the input challenge. With this improvement you can now use both Block and Mask Entry challenges, so sensitive information can be systematically perverted from user inputs before they reach FM.
This feature deals with the critical need for customers using enabubling applications for processing legitimate questions that could naturally contain PII elements without required to complete the requirement, providing greater flexibility while maintaining privacy. The ability is particularly valuable to applications where users can link to personal information in their questions, but still need safe and complicate answers.
New improving the handrail of functions
These improvises the functions of improvements across all police, making Amazon Bedrock railing more efficient and easier to implement.
Mandate -Amazon Bedrock Guarrails now implemented iam based on sweeping new bedrock:GuardrailIdentifier Key condition. This ability helps team security and compliance teams to create mandates for each model to make sure that the principles of organizational security are constantly subjected to all AI interactions. The status key can be used on InvokeModel,, InvokeModelWithResponseStream,, Converseand ConverseStream API. When the railing configured in IAM des policy does not correspond to a specific railing in the application, the system automatically rejects the application with the exception of rejected access and promotes compliance with organizational policies.
This centralized check will help you address critical challenges in the field of management, including the adequacy of content, security problems and privacy requirements. It also deals with a key challenge for the management of the company AI: to ensure that security checks consist in all AI interactions, and whose team or individuals develop these applications. Complex can be verified by understanding monitoring using the Amazon Cloudwatch or Amazon Simple Storage Service (Amazon S3) logs included routes documentation to show when and how the content was filtered.
For more information about this ability, see a detailed notice.
Optimize the performance of holding with a selective application for a handrail policy – Previously, Amazon Bedrock Guardrails applied the principles to inputs and outputs by default.
Now you have granugular control of the railing principles that you help you apply is selective to inputs, outputs or both – increase performance through targeted protection checks. This accuracy reduces unnecessary overhead processing costs and increases liability times while maintaining the necessary protection. Configure these optimized controls via amacon amonsons Bedrock API or Applicguardrails API for performance and safety balance according to your specific use requirements.
Analysis of policy before deployment for optimal configuration – The new monitor or analysis mode will help you evaluate the effectiveness of the warranty without direct application of the principles on the application. This ability allows a quick iteration to provide the visibility of how the configured railing would work, helping you experiment with different combinations and strengths of policies before deployment.
Get into production faster and safely with Amazon railing today
New capacity for Amazon Bedrock Guardrails APUT Our continuing commitment to help customers effectively implement AI on a scale. Detection of multimodal toxicity extends image protection, based on the establishment of IAM policy to responsible compliance, selective application of policy provides granular control, monitoring mode allows thorough testing before deployment and camouflage PII promotes promoting private private privacy. Together, these skills provide you with the tools you need to customize security measures, and Mainain consists of protection across your generative AI.
If you want to start with these new abilities, visit the Amazon Bedrock Console or Red on Amazon Bedrock Guardrails. More information about building responsible generative applications AI, Fer on AWS AI.
– Esra
Updated 8 April – Remove on the customer’s offer.
How’s the Blog of news? Take this 1 minute survey!
(This survey is hosted by an external company. AWS processes your information as described in the AWS Privacy Notice. AWS will own data collected via this survey and will not share information with dew survey.)