Hi OpenAI team,

we are seeing a read-only ChatGPT App MCP tool call being blocked by OpenAI safety checks before it reaches our MCP server.

The tool output is:

This tool call was blocked by OpenAI’s safety checks. Please double check what you are sending.

The trace shows:

{
  "externalCallTimeMs": null,
  "responseMetadata": null,
  "state": null
}

Because externalCallTimeMs is null and we see no corresponding server logs, this appears to be blocked before the MCP request is sent.

Example tool call

{
  "originIds": ["FRA", "DUS"],
  "destinationIds": ["IBZ", "PMI"],
  "departureMonths": ["2026-07-01", "2026-08-01"],
  "journeyType": "ONE_WAY",
  "passengers": {
    "adults": 1,
    "adolescents": 0,
    "children": 0,
    "infants": 0
  },
  "localeId": "en_US"
}

This is a benign read-only flight search.

We also tested equivalent payloads using destinationRegionId and alias fields like originAirports, destinationAirports, and travelMonths; all were blocked in the same way.

What we checked

• MCP endpoint authentication works.
• The server can be reached and tool definitions load.
• Direct calls through another MCP/client path work.
• The backend returns valid flight offer data for these routes/months.
• The tool does not create, update, delete, book, send, or mutate anything.
• Tool metadata is aligned with Apps SDK guidance:
  • readOnlyHint: true
  • destructiveHint: false
  • openWorldHint: false
  • _meta.securitySchemes: [{ "type": "noauth" }]
  • widget resources use text/html;profile=mcp-app

Question

Could you help identify why this read-only MCP tool call is being blocked before reaching the MCP server?

Is there a way to get more specific diagnostics for the safety-block reason, or is this possibly a false positive?

Bumping this for visibility. I am seeing this as well. Many benign queries are being blocked in my app as well. All of which were previously working fine.

Getting the same issue, was working fine yesterday

same issue here and it was working fine previously

I have also seen this fail since sometime yesterday. Absolutely benign calls are getting blocked. For example, I send this to my personal MCP server for GItea to retrieve a file:

{
"repo": "mcp-diagnostic-harness",
"path": "pyproject.toml",
"ref": "main"
}

This is the reply I get:

“This tool call was blocked by OpenAI’s safety checks. Please double check what you are sending.”

Additionally, I am seeing “Error in message stream” messages appear in the middle of the model replying lately as well.

Tom

Seeing the same issue as well!