I looked as well, but I found nothing comparable. Closest thing would be poe.com. The setup is more complicated though.
A human from OpenAI contacted me about 3 hours ago and is looking at this thread for reference and says they’re are working to resolve the issue. Seems like the agent update they made on Friday killed the Action functionality. I’ll update when I hear more.
7 Likes
Very glad I found this thread! I just happened to create my first action-powered GPT this weekend and can confirm that (even with API key using Basic auth - not just OAuth) this issue occurs. GETs are nice and quick even going through Ngrok…
2 Likes
Glad a “real human” (wow!) is looking at the issue 
And I don’t want to hijack this thread, but “my” problem has something to do with the same situation, I think.
It also started the 16th of August, while it was running stable for more than a year before.
And since “i discovered” the temporary work-a-round for the issue (“set always allow”), I hope someone will see this bug in the same time as the other one we are discussing here 
The thing is, get does work with the work-a-round, but it doesn’t remember it’s value it receives from the external server IF YOU ARE NOT USING IT IMMEDIATELY.
I have made a simpel test set-up, that demonstrates it;
- User enters a random city
- GPT connects to my server
- Server give a random response of two words
- GPT is told (in the instructions) to save both those words as “foobar”
- GPT is told to write down the first word immediately
- When user asks for second word after that 1st message, the GPT starts hallucinating
This is a very simple test, but with real (more complex) actions it doesn’t work as well (in both 4 and 5).
E.g. my server gives back a full article, GPT is said to write only the subject (it does) and ask me what to do next (with the raw article it received).
From the 2nd message on (after it did received the data) it doesn’t use/ can’t use the received data at all (since it did not write down the complete article in the 1st message).
I have tested it for about 40 hours now (debugging), several servers (on my side), GPT’s, solutions, environments and actions / schemes,
Nothing works since the exact same time as the “authorization bug” occurred.
Since Aug 14–16 KST, my Custom GPT Actions hang at “Talking to khsnfaraday.shop…” with no error message.
Key facts:
- Same OpenAPI schema worked perfectly until Aug 13.
- Server logs show zero requests at the exact timestamps.
- Repro with a minimal spec (only GET /health → returns {“ok”:true}, application/json).
- Also tested with a request-capture endpoint; no outgoing HTTP request from ChatGPT is observed.
- curl with the same URL/body/headers returns 200 OK immediately.
This indicates the client-side Action pipeline fails before sending any network request.
Please investigate the Actions execution path and error handling (UI shows indefinite spinner instead of an error).
We are on a Plus account (no enterprise domain restrictions).
Needed: a fix or guidance; at minimum an explicit error in UI.
cuz gpt 5 dont trust static domains with certs - use a tunneler like ngrok
We’re all waiting on feedback here:
8 posts were split to a new topic: Actions seem to be broken in the Custom GPT builder
I am sick of GPT releasing functions without testing properly.
current issue - GPT is not asking for user to provide approve access to the api, it just directly tries to connect ..saying user permission not received, doesn’t force user to approve or decline the connection, this started happening since 17th Aug 2025. I got this error from the test, on main GPT it just shows ‘talking to app….’
I have tried updating system prompt and asking them to force approval before connecting to api or after 1st fail, but it just goes on and on and does not give any feedback keep saying ‘talking to app….’ And does not connect to my webhook.
This is happening since 17th Aug 25
I have also been struggling with connecting my n8n workflow self hosted on hostinger VPS, since launch of gpt5, I could not solve it fully and then moved to n8n cloud setup, that worked fine and I was looking to solve this ‘error in connecting…’ issue, now this new issue has started happening since yesterday.
Don’t know where to go, can’t find anything on OpenAI portal, no way to contact them, I have 10 gpts none of them are working. very frustrating
I have tried deleting the action, adding new action, adding new webhook etc..
Can anyone help?
1 Like
I am facing the same issue. None of the Action APIs are working now which used to work seamlessly a couple of days back. Any workaround ??
2 Likes
You’re not crazy… Since GPT-5 release CustomGPT Actions have been broken. Everyone is on this thread awaiting answers from the OpenAI team. It is very reckless how they did the entire rollout.
1 Like
Yes, we are all in the same boat.
Work-a-round is setting privacy to always allow (for basic $_GET actions).
cuz gpt 5 dont trust static domains with certs - use a tunneler like ngrok
Has nothing to do with it.
1 Like
If changing Privacy settings → Actions → Always allow didn’t help in your case, try logging in through Privacy settings → Connected accounts.
That worked for me (though it only fixes basic GET actions for sure).
I tried that too, it did not work for me, interestingly it shows differently on different channels, desktop app shows the option, browser doesn’t. It doesn’t work.
The issue is still unresolved by OpenAI, and Actions remain broken.
It would be helpful if everyone reading or contributing to this thread could forward it to support@openai.com and also message OpenAI employees to draw their attention.
If we fail to raise sufficient awareness, based on my previous experience with how OpenAI handles custom GPT bugs, I’m afraid the issue could remain unfixed indefinitely - potentially for six months to a year or longer. That would be a nightmare scenario if it left unresolved.
2 Likes
i did it but no answer at all
1 Like
I don’t know if the following can help the community, but I’m posting it in case it might. So, like everyone here, my personal GPTs with “Action” stopped working three days ago and remained stuck. The trick of setting the privacy parameters to “always allow” worked for GPTs using GET, but not for those using POST.
So I gave the YAML file of my “Action” to ChatGPT and asked if it was possible to replace POST with GET. It suggested a test with Invoke-RestMethod in PowerShell, which consequently allows a direct call to my Endpoint. The test worked (my Endpoint is Make. com).
I then asked ChatGPT for the full YAML, duplicated my GPT, pasted the new YAML with GET, set the privacy parameters to “always allow”… and the GPT works just like before.
I understand that this probably won’t be the temporary solution for everyone, but maybe it can help some.
2 Likes
The issue persists for me. Using an ngrok gate and a handful of API endpoints.
GET endpoints usually work, but with very complex parameters are stuck in getting approval.
POST is stuck in getting approval.
Yep - same issue here today and our custom GPT is still failing/hanging despite the official status page suggesting it’s all fine, which it clearly isn’t. Have submitted a ticket to support as well
2 Likes