
Restricting AI Model Providers in AWS Bedrock: The Azure Policy Equivalent
Azure Policy lets you lock Azure AI Foundry down to approved model providers by name. Bedrock has no matching policy engine — but the same outcome is achievable through IAM, and the mechanism reveals something genuinely different about how each cloud enforces governance.
The question this starts with
If your organization runs Azure AI Foundry, you can restrict which
model providers are deployable using Azure Policy — models stay
visible in the catalog, but any attempt to deploy a disallowed one is
blocked before it happens. Specifically, Microsoft ships a built-in
policy, "[Preview]: Azure Machine Learning Deployments should only
use approved Registry Models," which supports allowlisting by
publisher — so restricting to Anthropic and OpenAI is a
configuration, not custom engineering. There's also a custom-policy
option with an "Allowed AI models" parameter for finer-grained control
by specific model name and version.
Move to Bedrock and the first instinct is to look for the same kind of
policy resource. There isn't one, because Bedrock models aren't
"deployed" the way an AI Foundry model deployment is — they're invoked
directly by model ID. That structural difference moves the enforcement
point from deployment time to call time, via IAM.
The AWS mechanism: IAM scoped to model ARNs
Every Bedrock foundation model has an ARN following a predictable
pattern:
arn:aws:bedrock:<region>::foundation-model/<provider>.<model-name>Restricting inference by model means denying or allowing based on
that ARN. Here's the deny pattern, using the full action list —
batch-inference jobs are easy to leave out and worth including:
{ "Version": "2012-10-17", "Statement": { "Sid": "DenyInference", "Effect": "Deny", "Action": [ "bedrock:InvokeModel", "bedrock:InvokeModelWithResponseStream", "bedrock:CreateModelInvocationJob" ], "Resource": "arn:aws:bedrock:*::foundation-model/model-id" }}For an allowlist rather than a denylist — restricting to specific
providers by ARN prefix — the same action list, inverted:
{ "Version": "2012-10-17", "Statement": [ { "Sid": "AllowOnlyApprovedProviders", "Effect": "Allow", "Action": [ "bedrock:InvokeModel", "bedrock:InvokeModelWithResponseStream", "bedrock:CreateModelInvocationJob" ], "Resource": [ "arn:aws:bedrock:*::foundation-model/anthropic.*", "arn:aws:bedrock:*::foundation-model/openai.*" ] } ]}Making it organization-wide: this same policy shape can be
attached as a service control policy (SCP) at the AWS Organizations or
OU level instead of to an individual role — the structural equivalent
of assigning Azure Policy above a set of subscriptions.
Five things worth knowing before you rely on this
1. Denying
InvokeModel blocks more than just InvokeModel .
Converse and StartAsyncInvoke are blocked automatically as soon as
InvokeModel is denied — you don't need to list them separately. If
you want to be explicit about it anyway, use bedrock:InvokeModel* as
a wildcard, or add bedrock:Converse and bedrock:ConverseStream to
the action list directly.2. Some models need two ARNs granted together, not one. Several
current-generation models route through a cross-region inference
profile rather than being invoked directly, and a policy covering
these needs both ARN types in the same
Resource array:"Resource": [ "arn:aws:bedrock:*::foundation-model/*", "arn:aws:bedrock:<region>:<account-id>:inference-profile/*"]Scope only the foundation-model ARN for one of these models and
invocation fails — test explicitly rather than assuming a policy that
looks complete actually covers what you're using.
3. Anthropic and OpenAI specifically carry an extra layer that
Amazon's own models don't. Models from Amazon, DeepSeek, Mistral AI,
and Meta aren't sold through AWS Marketplace and don't have Marketplace
product IDs. Anthropic and OpenAI models do — both are third-party
providers on Bedrock, and accessing their models runs through an
aws-marketplace:Subscribe permission, gated by the
aws-marketplace:ProductId condition key, on top of the standard
bedrock:InvokeModel restriction. For a policy restricting to exactly
these two providers, that's a second control surface worth locking
down alongside the ARN-based policy above, not instead of it.4. Whether model access is "on by default" varies by model, and it's
worth checking the current state for the specific ones you're
restricting rather than assuming. OpenAI's gpt-oss models are
auto-enabled with no request step. Anthropic's models involve an
explicit EULA-acceptance step tied to the Marketplace subscription
flow. Check the Bedrock console's model access page for your account
before assuming either behavior for a model you haven't tested.
5. This exact enforcement layer has failed silently before.
Independent security research found that the
aws-marketplace:ProductId
condition key didn't correctly restrict certain Bedrock models as
intended in production. AWS confirmed the issue and shipped a fix in
January 2024 — long resolved, but a fair reminder that a policy that
reads correctly and a policy that's actually enforced are two
different claims. Test a denied call directly, don't just review the
JSON.What to take away
If you're a student comparing the two clouds: the interesting part
isn't "can both clouds do this" — they both can. It's that Azure
enforces this at deployment time through a declarative policy engine,
while AWS enforces it at call time through IAM, with a second,
separate subscription-layer control specifically for third-party
providers. Same governance goal, genuinely different failure modes —
and worth knowing which one you're relying on before assuming a
written policy equals an enforced one.
If you're implementing this for real: don't stop at writing the
policy. Confirm a call to a non-approved provider is actually
rejected, confirm
Converse behaves as expected given point 1, check
whether your approved models need inference profile ARNs per point 2,
and if you're restricting to Anthropic or OpenAI specifically, cover
the Marketplace subscription layer from point 3 as well as the
InvokeModel layer — restricting one without the other leaves a real
gap, not a theoretical one.Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article