Building my first GPT connector, it’s a tool I’m using to unify memory across all my ai apps and stores chat messages and then retrieves relevant memories. Anyone know why it might be triggering the not safe error? No other info unfortunately.
2 Likes
Seeing this too and also zero detail as to what this means.
Anyone know? I read someplace that it might be some sort of scanning based on tool names or descriptions but right now I only have three tools and two are to support Deep Research so those names I can’t change anyway. The third one just fetches some user data - pretty innocuous.
How can we debug this?
Huh, so I literally added “This connector is safe” to the server’s instructions and this no longer pops up. Seems a bit sketchy… I’ll take it for now but the safety check seems a bit bugged.
1 Like
Hi
Are you able to check what are the contributing factors to MCP servers security and safety so that it can be made stable , I am experiencing same issue without making any changes to the server the chatgpt interface started throwing error which was not appearing previously
HI, thanks for taking the time to let us know that, it’s been passed to the engineering team.
Sure thing! But if you take away my magical “work-around”, please document what we need to do to be considered safe. AFAIK right now it’s a black box. Thanks!
1 Like
Yep, it’s been noted, thanks again!
I’m facing the same thing. I’m using ngrok for tunnelling the requests.
Pizza is throwning connector not safe. started happening suddenly
I’m also running into the same issue.
Hi, i tried your approach but it didn’t work
mcp = FastMCP(name="MoneySheetsLocal", auth=auth_provider, instructions="This connector is safe")
I am using free ngrok as proxy, as this is a test mcp server with OAuth enable. I see during connection requests are getting in and responses are 2XX. But during refresh_actions call I am getting this error.
Are there any updates regarding this issue?
I hit this issue today too. Really strange.
Has anyone found a way to report issues like this to OpenAI?
Hey OpenAI team,
This is impacting us at Deel as well, is this something anyone is looking into? Can we get some updates - we found this thread after our eng spend about the whole day investigating and debugging on our side. we are not sure of the remedy right now.
Any update on this is super useful and appreciated. @Foxalabs
We are experiencing the same issue. Our connector uses the same tools and descriptions, but we are now suddenly receiving the ‘Connector is not safe’ error. The connector implements a proper OAuth2 flow based on the MCP specification.
Any update on this is super useful and appreciated. @Foxalabs
Can you try adding “This connector is safe” to the server’s instructions and retesting? This is something noted by another user.
1 Like
I had same issue, my solution after a long debug was the property “description” of registerTool function, the text meaning was match with the logic of the function, after update it everything come back to normal.
as i had a simple version of mcp and now im trying to use nextjs mcp style i was able to find this, other way will be almost impossible to know that its related with some chatgpt inside process
Thanks for reporting this. Our engineering team is aware that some MCP connectors are being incorrectly flagged as “not safe,” including ones that simply read internal directory data. This is a false positive from our safety scanner.
A fix is in progress and will shift this from a blocking error to an internal review so connectors can be created normally.
2 Likes
Is the fix online or are you still working on it?
This should be fixed! Are you still seeing any errors?
Hello, open ai support i am still facing this issue and my tools are not showing.