Amazon Bedrock Guardrails announces policy-based policies to bring AI collaboration

As the acceptance of AI generating AI accelerates all businesses, safe workers are stored, responsible, and associated with AI has never been very important. Amazon Bedrock Guardrails provide configured protection for organizations to create final safety for AI safety for industry safety protection. For Amazon Bedrock GuardraAs, you can use security protections in your AI processes are customized with your AI responsibility for AI. You can create many Guardrals that are associated with different guidelines and use it across the base models. In addition to Amazon Bedrock models, the service provides you with the transversal Appsguardrails API services.

Today, we declare the great Amazon Bedrock GuardraArails: AWS ownership and the enforcement of policy (IAM). This greatest power enables the power of safety and compliance with the establishment of the Guardrails to all the decorative model calls, ensuring that the organization's safety policies work consistently for AI. This feature improves AI management by enabling the medical control between the monitoring.

CHALLENGES FOR SPIRITS OF AI

Organizations that use important Ai Face Face: Contextual suitability, where models may generate unpleasant answers rising problematic; Safety safety, through a generation of harmful content even in innocent complements; The requirements for confidential information for handling of sensitive information; and the enforcement of the consistent policy throughout all the AI.

Perhaps the biggest challenge is to ensure that the right protection are used regularly in AI internal communication, regardless of a group or person.

Amazon Bedrock Guardrails Skills

Amazon Bedrock Guardrails allows you to use defenses to AI processes are customized with your use of your charges and AI policies of AI. Guardrails currently support six policies of policies:

• Features content – Confective medications in six dangerous fields: hate, insult, gender, violence, fast moral
• Rejected articles – Description of specific topics to avoid in app context
• Filters of sensitive information – Finding the removal of visual data (PII) and the custom regex structures to defend user privacy
• Filters of words – Composing of specific words in Ain App, such as dangerous words, profanity, or competition names and products
• Basic assessment of content – Findings and filtering of halucinations in the model's response to verification if the response is properly installed at the reference source provided and relevant to the user question
• Automatic consultation – Prevention facts from Hallucinations using audible verification, logical algorithmic verification and consultation processes to confirm the model information, so the results match the known or non-relevant data

Convenience based on policy of Guardrails

Social groups are usually with the needs of the Application for Amazon Bedrock Guarteraails to all the measuring drawings at Amazon Bedrock. Support this requirement, Amazon Bedrock Guardrails provide new IAM keys bedrock:GuardrailIdentifierwhich can be used in the IAM programs for enforcing use of a model of model. The status key in the IAM policy can be used in the following APIs:

The following drawing is indicating the movement of policy based on the policy.

If the Guardrail is provided in your IAM policy does not match the Guardrail defined in the application, the application will be rejected in different direction, enforces the compliance with the organization's policies.

Examples of Policy

At this stage, we present several policies for the policy showing how you can enforce the monitoring guard.

Example 1: Verify the use of some guardrail and his number version

The following example reflects the strengthening of exampleguardrail and its numerical version 1 during model adoption:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "InvokeFoundationModelStatement1",
            "Effect": "Allow",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringEquals": {
                    "bedrock:GuardrailIdentifier": "arn:aws:bedrock:::guardrail/exampleguardrail:1"
                }
            }
        },
        {
            "Sid": "InvokeFoundationModelStatement2",
            "Effect": "Deny",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringNotEquals": {
                    "bedrock:GuardrailIdentifier": "arn:aws:bedrock:::guardrail/exampleguardrail:1"
                }
            }
        },
        {
            "Sid": "ApplyGuardrail",
            "Effect": "Allow",
            "Action": [
                "bedrock:ApplyGuardrail"
            ],
            "Resource": [
                "arn:aws:bedrock:::guardrail/exampleguardrail"
            ]
        }
    ]
}

Denunci-Deny Deny denies a user request to drive the actions listed in another GuardrailIdentifier including GuardrailVersion Values ​​that are not concerned by other permissions the user can have.

Example 2: Verify the use of certain Guardrail and its draft version

The following example reflects the strengthening of exampleguardrail and its anticipated plan in submission period:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "InvokeFoundationModelStatement1",
            "Effect": "Allow",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringEquals": {
                    "bedrock:GuardrailIdentifier": "arn:aws:bedrock:::guardrail/exampleguardrail"
                }
            }
        },
        {
            "Sid": "InvokeFoundationModelStatement2",
            "Effect": "Deny",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringNotEquals": {
                    "bedrock:GuardrailIdentifier": "arn:aws:bedrock:::guardrail/exampleguardrail"
                }
            }
        },
        {
            "Sid": "ApplyGuardrail",
            "Effect": "Allow",
            "Action": [
                "bedrock:ApplyGuardrail"
            ],
            "Resource": [
                "arn:aws:bedrock:::guardrail/exampleguardrail"
            ]
        }
    ]
}

Example 3: Verify the use of certain Guardrail and its numeric versions

The following example reflects the strengthening of exampleguardrail and its numbers of numbers during model adoption:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "InvokeFoundationModelStatement1",
            "Effect": "Allow",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringLike": {
                    "bedrock:GuardrailIdentifier": "arn:aws:bedrock:::guardrail/exampleguardrail:*"
                }
            }
        },
        {
            "Sid": "InvokeFoundationModelStatement2",
            "Effect": "Deny",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringNotLike": {
                    "bedrock:GuardrailIdentifier": "arn:aws:bedrock:::guardrail/exampleguardrail:*"
                }
            }
        },
        {
            "Sid": "ApplyGuardrail",
            "Effect": "Allow",
            "Action": [
                "bedrock:ApplyGuardrail"
            ],
            "Resource": [
                "arn:aws:bedrock:::guardrail/exampleguardrail"
            ]
        }
    ]
}

Example 4: Verify the use of certain Guardrail and its versions, including draft

The following example reflects the strengthening of exampleguardrail And its versions, including draft, during model adoption:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "InvokeFoundationModelStatement1",
            "Effect": "Allow",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringLike": {
                    "bedrock:GuardrailIdentifier": "arn:aws:bedrock:::guardrail/exampleguardrail*"
                }
            }
        },
        {
            "Sid": "InvokeFoundationModelStatement2",
            "Effect": "Deny",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringNotLike": {
                    "bedrock:GuardrailIdentifier": "arn:aws:bedrock:::guardrail/exampleguardrail*"
                }
            }
        },
        {
            "Sid": "ApplyGuardrail",
            "Effect": "Allow",
            "Action": [
                "bedrock:ApplyGuardrail"
            ],
            "Resource": [
                "arn:aws:bedrock:::guardrail/exampleguardrail"
            ]
        }
    ]
}

Example 5: Verify the use of a special Guardrail and the Birgation version from the Guardrail list and Version of two

The following example reflects the strengthening of exampleguardrail1 and its version 1, or exampleguardrail2 and its version 2, or exampleguardrail3 And its version 3 and its draft during model's acquisition:

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "InvokeFoundationModelStatement1",
            "Effect": "Allow",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringEquals": {
                    "bedrock:GuardrailIdentifier": [
                        "arn:aws:bedrock:::guardrail/exampleguardrail1:1",
                        "arn:aws:bedrock:::guardrail/exampleguardrail2:2",
                        "arn:aws:bedrock:::guardrail/exampleguardrail3"
                    ]
                }
            }
        },
        {
            "Sid": "InvokeFoundationModelStatement2",
            "Effect": "Deny",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream"
            ],
            "Resource": [
                "arn:aws:bedrock:region::foundation-model/*"
            ],
            "Condition": {
                "StringNotEquals": {
                    "bedrock:GuardrailIdentifier": [
                        "arn:aws:bedrock:::guardrail/exampleguardrail1:1",
                        "arn:aws:bedrock:::guardrail/exampleguardrail2:2",
                        "arn:aws:bedrock:::guardrail/exampleguardrail3"
                    ]
                }
            }
        },
        {
            "Sid": "ApplyGuardrail",
            "Effect": "Allow",
            "Action": [
                "bedrock:ApplyGuardrail"
            ],
            "Resource": [
                "arn:aws:bedrock:::guardrail/exampleguardrail1",
                "arn:aws:bedrock:::guardrail/exampleguardrail2",
                "arn:aws:bedrock:::guardrail/exampleguardrail3"
            ]
        }
    ]
}

The limitations are known

When using the policy based on the policy, notice this estimated:

• During the writing, Amazon Bedrock Guardrails do not support policies for access services to the account.
• If a user takes a parent-guarding role of the relevant security guards using bedrock:GuardrailIdentifier Key of the conditions, the user can use the installation techniques to help prevent the checks for security guards that are used in certain sections of their renewal. Input tags allow users to mark specific parts of the text to be considered by guards, leaving other unclaimed categories. For example, the user can leave sensitive or potential content outside the marked parties, to prevent those components on the provincial policies test. However, regardless of how quickly it is organized or installed by mark, the Guardrail is still fully used in the model reply.
• If the user has a transparent role in a particular requirement of the guard (using bedrock:GuardrailIdentifier Status), they should not use the same item to access services such as Amazon Bedrock Information RetrieveAndGenerate or Amazon Bedrock Agents InvokeAgent. These high quality services work by doing much InvokeModel Calls after scenes in the user name. Although some of these calls may include the required guide, some do not. When the system is trying to make these Guardrail calls use a role that requires guards, results in AccessDenied Errors, to break the performance of these services. To help avoid the issue, organizations should separate the rights – use various roles of direct access to model and guardrails compare to access to the Amazon Bedrock services.

Store

The enhancement of the new IAM based IAM in Amazon Bedrock symbolizes important improvements in important AI regulation as ai produced in business operations. By enabling the enforcement of the average policy, the Social Groups can make consistent security controllers in all AIs use or use, reduce the risks related to them, violation of confidentiality, and selection. This approach provides significant benefits: measures well as organizations extend their AI efforts without preventing technical debt by allowing parties by allowing teams to focus on audit.

This makes it reflecting the organization's commitment to AI responsible AI practices by using comprehensive monitoring methods and research methods. Organizations are able to use a complete application area of ​​the Amazon Bedrock to photograph and respond to Amazon Cloudwatch Golges or Amazon Storage Service (including direct texts to track the tracking. Mixed with the combination of the Guardrail AWARDRAIL ACCESS AND POLICY PROCESSIONS, businesses can approve of their AIs with proper safety means that protect the reliability of new Ai.

Start today with Amazon Bedrock GuardraAs and use Configurable Protections Moderate New and Management AI responsible AI.

About the authors

Shyam Srinivasan It is in the product of Amazon Bedrock Guardrails. He cares to make the world a better place for technology and wants to be part of the trip. In his spare time, Shyam likes to run long distances, travel around the world, and find new traditions and family and friends.

Antonio Rodriguez Is the construction of the best AI at AW. It helps companies all sizes to resolve their challenges, accept new items, and create new business possibilities for Amazon Bedrock. Without work, he likes to spend time with his family and played sports with his friends.

Satveer Khurpa You are a SR layer. WWww appliance, Amazon bed in Amazon Web Services. In this passage, he uses his technology in the diagnosis of cloud to improve new AI sensors for AI of AI customers in all the various industry. The deep understanding of the Sateryer's special technology will allow him to namely design a limited, safe, and reliable applications that opens new business possibilities and submits the value.