Amazon Bedrock model invocation logging for AIP-C01

Try 3 questions on this

Pick an answer to see why each option is right or wrong.

1. A bank uses an FM on Amazon Bedrock to draft replies to customer complaints. Model invocation logging delivers every prompt and response to an Amazon S3 bucket. A financial regulator requires the bank to keep these records for 7 years. During the 7 years, no user in the account, including administrators and the root user, can be able to delete or overwrite a log object. After 7 years, the bank must be able to delete the records. Which solution will meet these requirements?

   Right. In the game that’s Renown for your AI Safety, Security, and Governance building.

   Not that one: D is right. In the game your AI Safety, Security, and Governance building would lose a little Renown, and this question would come back later.

   Your answerA.Add an S3 Lifecycle rule that transitions the logs to S3 Glacier Deep Archive after 30 days and expires them after 7 years. Also enable S3 Versioning.A lifecycle rule moves and expires objects on schedule. It does nothing to stop someone deleting them earlier.Your answerB.Add a bucket policy that denies s3:DeleteObject and s3:DeleteObjectVersion to all principals. Allow only the security team's IAM role to change the bucket policy.A bucket policy can be edited or removed, and the root user can always change it, so it is not WORM storage.Your answerC.Enable S3 Versioning and MFA delete on the existing log bucket. Store the root user's MFA device in a safe that only the bank's security officer can open.MFA delete adds a step, but the root user with the MFA device can still delete object versions.The answerD.Create a bucket with S3 Object Lock enabled and a default retention period of 7 years in compliance mode. Point the model invocation logging configuration at the new bucket.S3 Object Lock in compliance mode is the WORM control. The S3 page says: "In compliance mode, a protected object version can't be overwritten or deleted by any user, including the root user in your AWS account."

   Show the full explanation and doc quotes

   It adds that the object's "retention mode can't be changed, and its retention period can't be shortened." A default retention period on the bucket applies the lock to every log object that invocation logging writes, and the lock ends when the retention period does, so the records can be deleted after 7 years.

   Written from docs.aws.amazon.com/AmazonS3/latest/userguide/object-lock.html

2. A company runs GenAI applications that call the Converse API in three AWS application accounts, all in us-east-1. The security team keeps all audit data in a central Amazon S3 bucket in a separate log archive account, also in us-east-1. The team must collect the full request data and response data of every model invocation from the three application accounts in the central bucket. The solution must not use custom code. Which combination of steps will meet these requirements? (Select TWO)

   Right. In the game that’s Renown for your AI Safety, Security, and Governance building.

   Not that one: D,E is right. In the game your AI Safety, Security, and Governance building would lose a little Renown, and this question would come back later.

   Your answerA.Add a statement to the central bucket's policy that allows s3:PutObject for the bedrock.amazonaws.com service principal from each of the application accounts.The bucket policy looks like the missing piece for cross-account delivery. The logging configuration still refuses a bucket that another account owns.Your answerB.In each application account, set the central bucket in the log archive account as the S3 destination of the account's model invocation logging configuration.Pointing every account at one central bucket is the obvious design, but invocation logging only delivers to a bucket that the same account owns in the same Region.Your answerC.Create an AWS Lambda function in each application account that reads each Converse request and response, and writes a copy of both to the central bucket.A Lambda function that copies requests and responses is custom code, which the solution must not use.Your answerD.In each application account, create an S3 bucket in us-east-1, and set that bucket as the S3 destination of the account's model invocation logging configuration.Your answerE.Enable versioning on each application account's log bucket and on the central bucket. Add a replication rule to each log bucket, and allow its replication in the central bucket's policy.

   Written from docs.aws.amazon.com/bedrock/latest/userguide/model-invocation-logging.html

3. A bank runs one customer assistant in its AWS account. The assistant calls the Amazon Bedrock Converse API with a guardrail that has a sensitive information filter. The filter blocks every prompt that contains a credit card number, and the user then receives the blocked message. Model invocation logging is turned on and delivers logs to Amazon S3. The bank uses these logs only for occasional troubleshooting. A new policy states that no credit card number that a user types may be stored in any Amazon Bedrock log. Which solution will meet these requirements?

   Right. In the game that’s Renown for your AI Safety, Security, and Governance building.

   Not that one: B is right. In the game your AI Safety, Security, and Governance building would lose a little Renown, and this question would come back later.

   Your answerA.Keep model invocation logging turned on, and encrypt the log bucket with a customer managed AWS KMS key. Allow only the IAM role of the security team to decrypt the log objects in the bucket.Encryption and a tight key policy protect the logs, but the policy says the numbers may not be stored at all, and they still are.The answerB.Turn off model invocation logging for the account in the Region. Troubleshoot from application logs that record only the request ID and the guardrail action.The guardrail components page (https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html) states: "All blocked content from the above policies will appear as plain text in Amazon Bedrock Model Invocation Logs, if you have enabled them."Your answerC.Keep model invocation logging turned on with its current settings. The guardrail blocks each prompt that contains a card number before the model runs, so the prompt is not written to the invocation logs.The user sees only the blocked message, but the invocation log still records the blocked content in plain text, so the card numbers are stored.Your answerD.Keep model invocation logging turned on, and turn on Amazon Macie automated sensitive data discovery for the log bucket. Delete each log object in which Macie finds a card number.Finding and deleting the objects afterwards means each card number sits in S3 until Macie reports it. The policy forbids storing it at all.

   Show the full explanation and doc quotes

   Blocking a prompt does not stop the logging of what was blocked. The same note gives the remedy: "You can disable Amazon Bedrock Invocation Logs if you do not want your blocked content to appear as plain text in the logs." The bank uses the logs only for troubleshooting and runs one assistant, so turning logging off meets the policy.

   Written from docs.aws.amazon.com/bedrock/latest/userguide/guardrails-components.html