[Bug]: Unable to apply guardrails to restrict "web_search_preview" tool usage for AzureOpenAI model in LiteLLM proxy

[MiladMt11]

What happened?

I'm trying to define a guardrail rule in my LiteLLM proxy config to deny usage of a specific tool (web_search_preview) for Azure OpenAI models.

According to the documentation, it should be possible to specify guardrails (tool_permission) in the config.yaml file.
However, no matter how I structure the configuration, the tool is still allowed.

I have tried multiple variations of defining the guardrail, even with regex and wildcards to accept all strings as the tool_name, but the tool is still allowed. below is an example of one of the variations of my config file.

config.yaml:

model_list:
   model_name: gpt-5-mini
    litellm_params:
      model: azure/responses/gpt-5-mini
      api_key: os.environ/AZURE_RESPONSES_OPENAI_API_KEY
      api_base: url
      api_version: version

   model_name: gpt-4o
    litellm_params:
      model: azure/responses/gpt-5.1
      api_key: os.environ/AZURE_RESPONSES_OPENAI_API_KEY
      api_base: url
      api_version: version


guardrails:
   guardrail_name: block_web_search
    litellm_params:
      guardrail: tool_permission
      mode: "pre_call"
      rules:
        - id: "block_web_search_tool"
          tool_name: "web_search_preview"
          decision: "deny"
      default_action: "allow"
      on_disallowed_action: "block"
      default_on: true

Does tool_name match the tool "type" field, or should it match something else?

Has anyone successfully restricted a specific tool for Azure OpenAI models? am I defining the guardrail for it correctly?

Environment:

Running in: LiteLLM Proxy Docker container

Model: Azure OpenAI (gpt-4o-mini)

Tool being restricted: web_search_preview

Relevant log output

Are you a ML Ops Team?

No

What LiteLLM version are you on ?

v1.80.0 OAS 3.1

Twitter / LinkedIn details

www.linkedin.com/in/milad-taghikhani-202b97200