This page lists every permission AURÄI’s AWS connector asks for and why, so you can review them before you connect. Shadow AI opens in January 2027.
1. What it finds
AI services in use (Amazon Bedrock models, guardrails, agents and knowledge bases, Bedrock AgentCore, SageMaker and Amazon Q Business), who has called AI services in the last 90 days, whether model-call logging is on, which roles can use AI, and secrets that look like AI provider keys, by name only.
2. How you give access
Who connects: An AWS administrator who can create IAM roles; for a whole organization, an administrator of the management account.
You launch our CloudFormation template in your own account. It creates one role that only your AURÄI workspace can use, with your workspace’s own external ID. The role can list and describe the resources below, and nothing else.
3. Every permission we ask for
These are every action the role allows. AURÄI never uses more, even if the role is later given more.
| IAM actions | What it is for |
|---|---|
| account: | Find which regions are enabled, so each one is checked |
| organizations: | See the accounts in your AWS organization |
| iam: | See which roles and users can use AI services, confirm our role is the one our template made, and name the account |
| bedrock: | Check whether Bedrock model-call logging is on |
| bedrock: | List the Bedrock models, guardrails, agents and knowledge bases you run |
| bedrock-agentcore: | List the agents you run on Bedrock AgentCore |
| sagemaker: | List SageMaker endpoints, domains and notebooks |
| qbusiness: | List Amazon Q Business applications |
| secretsmanager: | Find secrets that look like AI provider keys, by name only, never the value |
| cloudtrail: | See who called AI services in the last 90 days, from CloudTrail’s event history |
4. What we never ask for
- Reading S3 objects
- Reading secret values, Parameter Store values, or decrypting anything with KMS
- Reading database items or querying databases
- Reading Lambda function or container settings, or log contents
- Calling any AI model, agent, knowledge base or endpoint
5. How to revoke access
Delete the CloudFormation stack (or the organization stack in your management account). That removes the role, and access ends at once.
6. What happens to what it reads
A scan session lasts at most 24 hours. When it ends, what the scan collected and every finding worked out from it are deleted within the hour, and your AI register stays in a signed file that you keep. We keep a record of which systems each scan called and when, never what they returned. The Trust Center has the rest.
7. Other connectors
- Microsoft 365 and Entra ID20 read-only Microsoft Graph permissions, granted by an administrator.See every permission
- GitHubA GitHub App with eight read-only permissions, on the repositories you choose.See every permission
- OpenAI API platformAn admin key you create, limited to reading, which we only read with.See every permission
- Anthropic ConsoleAn admin key you create, which we only ever read with.See every permission
- DNSPublic records only. Nothing to grant and nothing to revoke.See every permission