Web_search filters.blocked_domains appears to be ignored in some cases

API Bugs
1

Hi everyone,

I am using the Responses API with web_search tool and filters.blocked_domains.

Environment:

openai-python==2.38.0
model=gpt-5.4-mini

Request shape:

response = client.responses.create(
    model="gpt-5.4-mini",
    tools=[{
        "type": "web_search",
        "filters": {
            "blocked_domains": [
                "blocked-domain-1.com",
                "blocked-domain-2.com"
            ]
        }
    }],
    tool_choice="required",
    include=["web_search_call.action.sources"],
    input="Give me some basic information about the company with this website: www.test.com",
)

Expected behavior: blocked domains should not appear in web_search_call.action.sources.

Actual behavior: across 213 test cases, blocked domains appeared in sources in 13 cases.

I confirmed:

  • using web_search, not web_search_preview

  • using Responses API, not Chat Completions

  • domains are normalized, without https:// or paths

  • fewer than 100 blocked domains

One related observation: in openai-python==2.38.0, the generated SDK type file appears to include allowed_domains but not blocked_domains:

src/openai/types/responses/web_search_tool_param.py

So I am wondering whether this is an SDK typing/schema mismatch, a request serialization issue, or a backend enforcement issue.

Is this expected, or is blocked_domains enforcement currently unreliable?

Thanks.

2

Interesting report. The intermittent leakage rate (13/213) makes this seem less like outright unsupported behavior and more like either:

  • backend-side enforcement inconsistency

  • hostname normalization mismatch

  • or filtering occurring post-retrieval rather than pre-retrieval.

And one thing that may help isolate it further:

  • so compare SDK vs raw REST requests

  • and test redirect/canonical URL handling

  • and try testing subdomains separately (example.com vs www.example.com)

  • or inspect whether leaked domains only appear in sources or also influence generated output

If raw REST reproduces the same behavior, that would probably rule out SDK serialization and point more toward backend enforcement logic… maybe.. hmm… try

3

I have followed the proper rules by using domains such as example.com. However, the leakage of blocked domains seems to vary between test runs. Sometimes it is 5, sometimes 13, and sometimes 65. At this point, I am no longer sure what exactly web_search_call.action.sources contains, or whether it includes blocked domains.

4

Hi!

I was unable to reproduce this. Are you still experiencing these issues?

However, I can confirm a Python SDK typing/schema gap where blocked_domains is valid and accepted by the API but missing from the generated WebSearchToolParam type.

5

Hey, No, I haven’t encountered the issue again. However, I’ve only run the same workflow 2–3 times since I posted here. I’ll update this thread if I experience the issue again.