AI services guardrails limit models, keys and networks, and the library holds 19 of them
Facts read 10 Nov 2026. Page reviewed 16 Nov 2026.
19 guardrails in the library address ai services: 7 AWS, 7 Azure, 5 Google Cloud. Each entry gives the policy text, what it blocks, what it is known to break and how to test it first. Every score on this page was read from the evidence on 10 Nov 2026 and links to its source.
All risk classes ยท Download the library
Guardrails in this class govern model access, API keys, public network access and tools for AI services. The library holds 19, the largest class: 7 for AWS, 7 for Azure and 5 for Google Cloud. They break the work of the teams using the services. The AWS Bedrock allow list stops anyone invoking a model not matched by the approved entry, and denying Bedrock API keys stops existing keys and bearer tokens. Azure notes Azure OpenAI Studio requires key access and will not function if it is disabled. Google says enforcing its Grounding with Google Search constraint removes that feature from applications that use it.
What these guardrails break
- Region deny that allows Bedrock cross-Region inference (AWS): The placeholders {example-region-1}, {example-region-2} and the inference profile prefix must be replaced, and the README notes model access still needs the separate model approval SCP.
- Allow only approved Amazon Bedrock models (AWS): Users cannot invoke or subscribe to any model not matched by the approved entry; the placeholders must be replaced with a provider or model identifier before use.
- Deny Amazon Bedrock API keys (AWS): Existing Bedrock API keys and short-term bearer tokens stop working for every principal in the affected accounts, because the second statement has no exception condition.
- Limit Amazon Bedrock long-term API keys to under 30 days (AWS): Creating a Bedrock long-term API key with a validity of 30 days or longer fails for every principal in the affected accounts.
- Deny IAM and account actions invoked from chat applications (AWS): Requests for these IAM and Organizations actions made from chat channels fail for every principal; the same actions made from the console or CLI are not matched by the condition.
- Enforce an Amazon Bedrock guardrail on model inference (declarative policy) (AWS): The guardrail must be owned by the management account, must have a non-DRAFT version, and needs a resource-based policy allowing organization members to call ApplyGuardrail.
- Opt out of AI service content use organization-wide (declarative policy) (AWS): AWS documents that opting out deletes the associated historical content shared with AWS before the setting was applied, other than content needed to provide the service.
- Configure Cognitive Services accounts to disable public network access (Azure): Modify changes the property during create or update instead of rejecting the request, and existing accounts change only through a remediation task.
- Configure Cognitive Services accounts to disable local authentication (Azure): Modify changes the property during create or update instead of rejecting the request, and existing accounts change only through a remediation task.
- Configure Azure AI Services resources to disable local key access (Azure): Azure OpenAI Studio requires key access and will not function if key access is disabled, per the definition description.
- Configure Azure Machine Learning workspaces to disable public network access (Azure): Modify changes the property during create or update instead of rejecting the request, and existing workspaces change only through a remediation task; Microsoft suggests private endpoints to control exposure instead.
- Azure Machine Learning registry deployments restricted to the allowed registry (preview) (Azure): This definition is a preview, and if no allowed registry name is set the parameter description says all models from all registries are restricted.
- Microsoft Foundry agents and toolboxes should not use denied tools (preview) (Azure): This definition is a preview, and the same definition supports Audit, so a deny only applies when you set the effect to Deny.
- Azure AI Search services should disable public network access (Azure): Existing search services that are public are marked non-compliant but not changed; Microsoft suggests private endpoints to limit exposure.
- Restrict public IP access on new Vertex AI Workbench instances (Google Cloud): Google scopes the constraint to newly created notebooks and instances, and states that new organization policies are usually not retroactive.
- Disable root access on new Vertex AI Workbench notebooks and instances (Google Cloud): Google scopes the constraint to newly created notebooks and instances, and states that new organization policies are usually not retroactive.
- Define access to models on Vertex AI (Google Cloud): Google documents that the Model Garden policy only applies to models in Model Garden, not models registered in the Model Registry, and that a policy cannot exceed 500 allowed and denied values.
- Disable Grounding with Google Search in generative AI APIs (Google Cloud): Google documents that the feature is enabled by default, so enforcing the constraint removes grounding with Google Search from applications that use it.
- Restrict egress domains for data connectors (Google Cloud): Google documents that, when enforced, only URLs with an allowed FQDN can be used for data connectors in the covered projects.
Where Act is behind
Act writes guardrails in the provider's own mechanisms and hands them to your pipeline to enforce, so it is bounded by what those mechanisms can express in each cloud. It does not apply them itself, the boundary types it names are AWS's and Azure's, and exception and waiver handling after a guardrail ships was described as roadmap.
Related pages
Sources
- Region deny that allows Bedrock cross-Region inference, read 10 Nov 2026
- Allow only approved Amazon Bedrock models, read 10 Nov 2026
- Deny Amazon Bedrock API keys, read 10 Nov 2026
- Limit Amazon Bedrock long-term API keys to under 30 days, read 10 Nov 2026
- Deny IAM and account actions invoked from chat applications, read 10 Nov 2026
- Enforce an Amazon Bedrock guardrail on model inference (declarative policy), read 10 Nov 2026
- Opt out of AI service content use organization-wide (declarative policy), read 10 Nov 2026
- Configure Cognitive Services accounts to disable public network access, read 10 Nov 2026
Last reviewed 16 Nov 2026.