fix(terraform): reject wildcard accounts in bedrock_model_arns

The variable already documented that wildcards in the account segment are
not accepted, and the validation's own error_message said they are rejected,
but the regex used [^:]* there, so arn:aws:bedrock:*:*:imported-model/* passed
and was copied verbatim into the task role's policy. Require the account
segment to be empty -- the canonical form for AWS-owned foundation models,
arn:aws:bedrock:*::foundation-model/anthropic.* -- or a literal 12-digit
account id, which is what every inference-profile, imported-model and
provisioned-model ARN carries.

Two runs pin both sides: a wildcard account is rejected, and a concrete
account in a cross-region inference-profile ARN still passes, so the
tightening cannot quietly cost the form operators actually paste.

terraform/litellm/aws/README.md

@@ -220,9 +220,13 @@ foundation model in whichever region has capacity, so it needs both ARNs above:
 the profile itself, and the foundation model wildcarded across every region the
 profile can route to
 
-Every entry must be a Bedrock ARN. A bare `"*"` is rejected at plan time, since
-it would let the proxy's task role invoke every Bedrock resource in the account,
-including other teams' provisioned throughput and private imported models
+Every entry must be a Bedrock ARN naming a resource type, with the account
+segment either empty (foundation models, which AWS owns) or a literal 12-digit
+account id. A bare `"*"`, and any wildcard in the account or resource-type
+segment, is rejected at plan time, since it would let the proxy's task role
+invoke every Bedrock resource in the account, including other teams'
+provisioned throughput and private imported models. Wildcards inside the
+resource segment are what scope a family of models, and stay allowed
 
 Two adjacent flags cover the rest of the Bedrock surface. `enable_bedrock_mantle`
 grants `bedrock-mantle:CreateInference`, which the `bedrock_mantle/...` models

terraform/litellm/aws/examples/default/terraform.tfvars.example

@@ -139,8 +139,10 @@ env    = "stage"
 # reference them in proxy_config as `bedrock/<model id>`. A cross-region
 # inference profile needs both the profile ARN and the foundation-model ARN.
 # Every entry must name a Bedrock resource type ("foundation-model/...",
-# "inference-profile/...", "imported-model/..."); "*", "arn:aws:bedrock:*",
-# and "arn:aws:bedrock:*:*:*" are rejected at plan time.
+# "inference-profile/...", "imported-model/...") and carry either an empty
+# account segment or a literal 12-digit account id; "*", "arn:aws:bedrock:*",
+# "arn:aws:bedrock:*:*:*" and "arn:aws:bedrock:*:*:imported-model/*" are
+# rejected at plan time.
 # bedrock_model_arns = [
 #   "arn:aws:bedrock:*:111122223333:inference-profile/us.anthropic.*",
 #   "arn:aws:bedrock:*::foundation-model/anthropic.*",

terraform/litellm/aws/tests/bedrock.tftest.hcl

@@ -180,6 +180,37 @@ run "an_all_segments_wildcard_bedrock_arn_is_rejected" {
   ]
 }
 
+# A wildcard in the account segment reads like a scoped ARN but matches every
+# account the identity can reach, which is why the variable documents the account
+# as "empty or a concrete id". The resource-type slug clause above does not catch
+# it, so it needs its own guard and its own case.
+run "an_account_wildcard_bedrock_arn_is_rejected" {
+  command = plan
+
+  variables {
+    bedrock_model_arns = ["arn:aws:bedrock:us-east-1:*:imported-model/*"]
+  }
+
+  expect_failures = [
+    var.bedrock_model_arns,
+  ]
+}
+
+# The mirror of the case above: tightening the account segment must not cost the
+# ARN form operators actually paste for a cross-region inference profile.
+run "a_concrete_account_inference_profile_arn_is_accepted" {
+  command = plan
+
+  variables {
+    bedrock_model_arns = ["arn:aws:bedrock:*:111122223333:inference-profile/us.anthropic.*"]
+  }
+
+  assert {
+    condition     = length(aws_iam_policy.bedrock_invoke) == 1
+    error_message = "A 12-digit account in an inference-profile ARN must stay valid."
+  }
+}
+
 run "a_china_partition_bedrock_arn_is_accepted" {
   command = plan
 

terraform/litellm/aws/variables.tf

@@ -591,8 +591,8 @@ variable "bedrock_model_arns" {
   default     = []
 
   validation {
-    condition     = alltrue([for a in var.bedrock_model_arns : can(regex("^arn:aws[a-z-]*:bedrock:[^:]*:[^:]*:[a-z-]+/", a))])
-    error_message = "Every bedrock_model_arns entry must be a Bedrock ARN with a real resource type, e.g. arn:aws:bedrock:*::foundation-model/anthropic.*. Bare wildcards on the account or resource-type segment (\"*\", \"arn:aws:bedrock:*\", \"arn:aws:bedrock:*:*:*\") are rejected, because they would let the task role invoke every Bedrock resource in the account, including other teams' provisioned throughput and private imported models."
+    condition     = alltrue([for a in var.bedrock_model_arns : can(regex("^arn:aws[a-z-]*:bedrock:[^:]*:([0-9]{12})?:[a-z-]+/", a))])
+    error_message = "Every bedrock_model_arns entry must be a Bedrock ARN with a real resource type, e.g. arn:aws:bedrock:*::foundation-model/anthropic.*. The account segment must be empty (foundation models) or a literal 12-digit account id; wildcards on the account or resource-type segment (\"*\", \"arn:aws:bedrock:*\", \"arn:aws:bedrock:*:*:*\", \"arn:aws:bedrock:*:*:imported-model/*\") are rejected, because they would let the task role invoke every Bedrock resource in the account, including other teams' provisioned throughput and private imported models."
   }
 }