Open ai safety checks blocking benign file writes to google drive folders

[kernowspace]

Intermittent “tool call was blocked by OpenAI’s safety checks” on Google Drive — benign writes blocked before Drive receives them

I’m trying to establish whether this is affecting other ChatGPT users or just my account/workflow.

I use the Google Drive connector as the transport layer for a personal local workflow. It had been working normally for a long time. On 12 September 2026 I began seeing intermittent pre-dispatch refusals on completely benign Drive actions.

The exact error is:

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

The important point is that the blocked requests never reach Google Drive. No destination file is created and my downstream system receives nothing.

I can reproduce both success and failure from the same ChatGPT account and the same Drive connection.

• 12 Sep, 23:34 UTC: harmless read-only request to read only /etc/os-release was accepted, written to Drive, processed downstream and completed successfully.

• 12 Sep, 23:36 UTC: a very similar read-only control was blocked at the prerequisite Drive register write before the actual job could even be submitted.

• 13 Sep, 13:42 UTC: simple desktop notification/tone test — blocked at the Drive register write.

• 13 Sep, 13:42 UTC: recurring five-minute timer registration — blocked the same way.

• 13 Sep, 13:42 UTC: narrowly scoped request to read one exact local file and append Hello world to one exact Desktop text file — again blocked at the Drive register write. It never reached the PC.

Earlier that day and on previous days, similar Drive operations had been accepted normally. I have also had a benign document successfully created in Drive and then had attempts to place that same document in another ordinary Drive folder blocked.

This doesn’t look like a conventional Google Drive permission issue because the same account/connector sometimes succeeds, the refusal happens before Drive receives anything, successful requests operate normally downstream, and even harmless audit/register writes can be stopped.

I’ve already reported this privately to OpenAI Support with screenshots and fresh reproductions. I received an initial reply but have had no response to two subsequent follow-ups.

I also found the OpenAI Developer Community thread where OpenAI Support acknowledged incorrect safety blocks affecting legitimate/read-only tool calls.

Is anyone else currently seeing this with Google Drive, GitHub, Notion, MCP tools or other ChatGPT connectors/apps?

If you are, it would be useful to compare the approximate time/date, connector used, exact error text, whether the destination service actually received the request, and whether retrying the same operation later succeeded.

I have a sanitized evidence packet containing successful and blocked examples and redacted JSON payloads if it would help compare cases.