Can confirm having the same issue for over 24 hours now. My custom gpt is for personal use and it’s frustrating enough - I can’t even imagine using it for business use if the radio silence from open AI is the norm on such things. Wild that this is the leading company in the field…
2 Likes
The issue still persists. Could someone from OpenAI please fix this? This is a critical issue!
4 Likes
This is so disheartening! I had high hopes for the reliability of this platform, but the lack of acknowledgement from OpenAI about an issue so crucial for users monetizing CustomGPTs is SCARY!!! I’ve lost at least 40 hours of time troubleshooting this since the release of GPT-5 thinking that I misconfigured something (even though I didn’t change anything from when it was working fine) as the issue has been sporadic… I’m deeply concerned about moving forward because I’d be issuing refunds to my users on a mass scale. Thank you guys for starting and contributing to this thread bc I was beginning to think no one was paying attention at all to any of this. I’m at a stand still awaiting a resolution from OpenAI, which seems not to notice even though I’m pleading to the support email to look into it. It keeps coming back with “Everything is fine, no outages.” This is a prime example where we need to keep REAL PEOPLE in key positions of oversight, such as this!!!
4 Likes
For $_GET it works.
But on mobile (Android app at least) it’s broke for like months.
When I change my custom GTP, Android app never asks for permission - it simple hangs on ‘contacting server’.
So I have to set it on mobile to “always allow” because the “allow button” never kicks in.
This is the case for at least 2 or 3 months.
1 Like
$_GET works, but in my case it doesn’t remember the data it gets.
I can see from both my server-side as in GPT debug environment that it DOES get the data requested.
But when I refer to it after the 1st message, it doesn’t remember it and starts hallucinating (it’s not using the data after the 1st message).
SCHEME
- Get data from server
- Receive data
- “do something with data” ← this works
- Get data from server
- Receive data
- GPT says :
data fetched what’s next - “do something with data” ← this fails
So when you have one single message between “receive data” and “do something with it” it doesn’t remember the data.
At least with my app (fetched an article from my server, has to write something about it).
I always use one message after ‘receive data’ to instruct hem what to write, but then it doesn’t remember it. So now I have to inform him what to write BEFORE it fetches the data at all.
It’s really broken. I have some-x custom GPT since 2023, all were stable and working with custom actions. 1000’s and 1000’s times. But suddenly it starts hallucinating (GPT 4 and 5).
I use it for business, but am only debugging the bugs from OpenAI most of the time. Even the @mentions inside a custom GPT are not working for a month now.
1 Like
I just tested Actions with OAuth using both POST and GET requests, and in both cases Actions with OAuth did not work. This issue occurs regardless of whether the Privacy setting for the Action with OAuth is set to “Always allow” or “Ask.”
Below are screenshot examples of an Action with OAuth using a GET request when the Privacy setting is set to “Always allow”:
2 Likes
In my case, GET is hallucinating on “always allow”, not works on “ask”. I’m saying it’s hallucinating because the server logs showing no requests from GPT at all. When I force the GPT to explain raw response it admits it simulated the GET response.
1 Like
Same thing happens here. Friday it was working fine, today it does not show the permission button anymore and gets stuck.
2 Likes
I don’t know what’s wrong.
Like I said, I am working with GPT’s from about 2023 and GET did work.
Since about 48 hours it’s not using any data if got from GET.
I do GET the data, because I have set it to “always allow”.
And I can see the GET call from GPT incoming to my server (it notifies me when GPT makes a call to it’s data).
The server sends the requested data and GPT does receive it (I can see that when using the [debug view] while customizing the GPT).
But when there is a single message between “getting” the data and “doing something with it” (even a . dot / period is enough), it will start hallucinating.
When I force it to use the data 100% direct after receiving it, it works.
But in my workflow, that’s not working for me (because I have to message GPT at least one time after it receives the data, to finetune what he has to do with it).
So, also this is broken.
1 Like
If someone from OpenAI is reading this, please fix it! This is a huge issue, and it’s not the kind of problem that can be left unresolved.
9 Likes
Hi,
I can confirm I’m experiencing the exact same issue. Everything was working perfectly until two or three days ago, and since then:
- GET requests still work fine (even with authentication).
- But all POST requests with authentication are now failing.
Some details:
- My own APIs using BasicAuth used to work flawlessly. Now, POST requests don’t even reach my server (no trace in the logs).
- When tested outside of ChatGPT (via curl or wget), the exact same POST + Auth requests work perfectly.
- Strange example: the Telegram API still works with POST with custom GPT (because the token is passed in the URL, so there is no explicit header-based authentication).
In short, the problem is specifically with the POST + Auth combination, not POST in general, and not tied to a particular domain.
Has anyone seen any update (here on the forum, via support, or elsewhere) about this bug or its planned resolution?
Thanks in advance if you have any news.
1 Like
Since August 16, all POST/PUT/DELETE requests from custom GPTs stopped working.
GET still works, but everything else just hangs on “Talking to domain…”. In Postman the same requests succeed.
This happened right as OpenAI announced the “calendar & email revolution” with GPT-5.
But in reality:
GPT can’t create events directly — only gives you a pre-filled form to confirm.
Reсurring events? Just an .ics file for import.
Real automation is gone.
For developers, this is devastating. We built real systems around custom GPT actions, and overnight they’re dead — with no explanation from OpenAI.
This isn’t a minor bug. It’s a serious blow to trust.
4 Likes
This is SO crazy. I’ve built my whole business around the customGPTs I spend all summer building that rely on POST calls. My GPT that relies on GET only still works, but I’m scrambling now to see if there is a way to fix my other GPTs so they work with GET instead of POST. It’s WILD OpenAI would change something that would break this functionality that is so commonly used in this community of custom GPT developers. Wow.
5 Likes
And it seems as though they don’t care…
@edesvigne
In my experience, GET works only with Actions that have no OAuth and only when ‘Always allow’ is selected in the Privacy settings. All request methods, including GET, fail when the Action uses OAuth.
Did I understand correctly that in your experience GET requests work when Action uses Authentication type “API Key” with Auth type “Basic” instead of Authentication type “OAuth”? Could you also check whether ‘Always allow’ was enabled in your case when using GET requests, and whether it still works if ‘Ask’ is chosen in the Privacy settings instead of ‘Always allow’?
I’ve also noticed that if ‘Always allow’ is not selected, all Actions fail to work, regardless of the authentication type or request method. First-time users of GPTs usually do not have ‘Always allow’ enabled, so for them, all GPTs’ Actions appear to be non-functional.
And when Actions use the OAuth authentication type, all request methods fail to work regardless.
”Has anyone seen any update (here on the forum, via support, or elsewhere) about this bug or its planned resolution?”
It looks like the only chance of this being fixed is if somebody from OpenAI pays attention to this thread and our messages.
2 Likes
GET works for me, but most likely only because I’m using an API-Key instead of OAuth.
The difference is that with an API-Key you can set the permissions as “permanent”, so GET requests go through without any extra confirmation.
However, all POST/PUT/DELETE requests — even with an API-Key — still require an action confirmation. Since the confirmation window no longer appears, those requests just hang.
With OAuth it’s even worse, because you need to fetch a token first, and that step also depends on confirmation. That’s why for many people even GET doesn’t work.
3 Likes
I’ve passed the link to this thread on to the support@openai.com bot and requested human intervention, so I’m supposed to be contacted in the next 2-3 days by a human and I’ll make these EXTREMELY CRITICAL issues made known to the oblivious staff. Hopefully it’s resolved by then, but worst case I’ll be following up here once I’m contacted.
2 Likes
If anyone wants to comment on my X.com post about this issue - maybe we can raise the issue publicly and get more attention - https://x.com/ranchoscorpio/status/1957180585887076528?t=ovcmzf96jhLo5Vn-GSeHLQ&s=19
3 Likes
I got this from the support email bot - not exactly helpful -
Support Team
| 3:58 PM (0 minutes ago) | ||
|---|---|---|
| undefined | ---- |
| to me |  |
|---|
Thank you for the detailed follow-up, Kathryn. Based on your description, this does appear to align with a recent, broader regression affecting Custom GPT Actions—especially POST requests and the consent workflow. Other developers reporting the same sudden breakage and changes in user permissions supports that this isn’t isolated to your environment.
From the available documentation and recent policy changes:
- There is no official statement indicating that POST Actions should be deliberately disabled by recent privacy updates.
- Actions (including POSTs) are still intended to work, with responses expected within 2 minutes, or they will time out.
- Newer workspace privacy controls now impact when (and if) users see permission prompts for external API connections, but these changes should not “silently” break POST Actions.
Key points from the Help Center:
- If API domains weren’t properly allowlisted by workspace admins (more common in Enterprise, but users who enable privacy controls may be affected too), Actions may silently fail.
- For GET-based Actions, user “Always Allow” privacy settings are now required, which is a new friction point.
- No recent official guidance suggests that POST calls, in particular, should simply stop working after a privacy update.
Put simply: There is no documentation or policy stating that POST Actions are now intentionally disabled, nor that the consent workflow should no longer appear. This matches a possible system-level or privacy-related bug introduced in the latest update.
Next steps:
- If you have not already, double-check that the relevant domains are allowlisted in your GPT’s admin settings. However, since your product worked prior to the update and GET calls are allowed, it’s highly likely the problem is not with your configuration.
- Please share any log snippets, OpenAPI schema samples, and payloads (or clear examples) if possible—these can help further.
- OpenAI engineering is aware that Actions should not fail silently, and the lack of proper prompts/consent signals unintended behavior.
I am making sure your feedback and these precise details are escalated to our technical team for urgent investigation, as this is clearly business-critical and affecting many users.
If there is an official update or workaround, I will reply here immediately. In the meantime, closely watch the OpenAI Status Page for urgent platform announcements or bug fixes. Thank you for your patience and rapid troubleshooting as we get this resolved.
1 Like