September 29 I was using ChatGPT's GitHub connector to work on repos like I had been doing already.
Nothing especially exotic. Create stuff, edit stuff, have ChatGPT operate against GitHub through the connector.
Then GitHub suspended my account.
Not a repo. The whole account.
That meant several FOSS projects I maintain just started returning 404 for a while. Actions were gone too, so anything depending on GitHub Actions or Pages was broken along with it.
Pretty fucking annoying way to find out an integration has had a bad day.
what i could actually see
I pulled the GitHub audit data afterward because the initial explanation I got didn't line up very well with what I was seeing.
The useful timestamps were close together.
I had been doing repo work through ChatGPT.
At about 22:27 I connected/configured another repo through the GitHub integration.
At 23:03:29 GitHub logged:
user.suspendThen:
user.unsuspendGitHub also revoked the OAuth tokens belonging to the ChatGPT connector. More than once.
So something clearly happened around that integration session.
I can't prove from the audit log alone that ChatGPT caused the suspension. GitHub has internal abuse signals I don't get to see and OpenAI has connector logs I don't get to see.
But I also couldn't find the usual obvious "somebody stole my GitHub account" shit.
No mystery foreign login.
No SSH key I didn't recognize.
No weird OAuth app.
No obvious malicious pushes.
The activity I could see looked like my activity.
Which left me staring pretty hard at the connector activity immediately before GitHub's abuse system fired and then nuked the connector's OAuth authorization.
github support
GitHub support mostly seemed to think I was reporting some kind of account compromise or suspicious login.
I kept trying to explain that this didn't fit what I was seeing.
The audit log showed GitHub suspending the account right after a bunch of connector activity and then GitHub itself revoking the ChatGPT OAuth credentials.
At one point the conversation also seemed to be operating on the assumption that the ChatGPT GitHub integration couldn't really be doing write operations.
Except I had literally been using it to do write operations.
So that got weird fast.
I don't think I ever managed to get GitHub support onto the exact question I was asking (dont know why im surprised, its a microsoft product):
what caused the suspension at 23:03, and why were the ChatGPT connector credentials revoked at the same time?
Maybe they understood it internally and just couldn't tell me anything useful. Maybe the people answering didn't have access to whatever abuse system made the decision. I don't know.
From my side it kept feeling like I was reporting one incident and getting help for a slightly different one.
Meanwhile repos for several open source projects I maintain had been 404 because GitHub had suspended the account hosting them.
openai support was weirdly good
I contacted OpenAI too because obviously I wanted somebody on their side looking at the connector.
I was expecting this to be painful.
It really wasn't.
The automated support side couldn't inspect the connector telemetry I was asking about, which is reasonable. I wasn't asking how to click "Reconnect GitHub." I wanted somebody who could look at what the connector had actually been doing around the time GitHub suspended me.
It got escalated.
I was talking to a human in less than 12 hours.
And the human response actually seemed to understand the issue I was reporting.
I wasn't asking them to restore GitHub. OpenAI can't do that.
I wanted them to look at their side of the integration and figure out whether it had gone nuts, retried something a stupid number of times, generated some bizarre API pattern, or otherwise done something that GitHub might reasonably have interpreted as abuse.
I have complained about plenty of OpenAI shit before and probably will again, but support handled this one well.
GitHub was the company that owned the account that got suspended, and somehow OpenAI was the side where I felt like a human actually understood what I was asking.
That was not the outcome I expected.
the connector is the interesting part
This is why I'm writing this up instead of just chalking it up to "GitHub false positive, whatever."
These integrations have real write access.
They can create repos, edit files, push changes, and generally act on your behalf.
If one of them starts behaving strangely and trips an abuse system, the blast radius isn't limited to whatever repo you were working on.
In my case several unrelated public projects disappeared for a while because GitHub suspended the entire account.
That's kind of fucking nuts.
Maybe ChatGPT retried something too aggressively.
Maybe GitHub has an abuse heuristic that doesn't play nicely with AI agents doing bursts of API operations.
Maybe something else happened that neither company's user-facing logs expose.
I still don't know.
What I have is the timing, the suspension event, the OAuth revocations, and the fact that nothing else in the visible account activity looked like an obvious compromise.
That's where the evidence stops.
after
The account came back.
The ChatGPT authorization did not magically come back with it because GitHub had revoked it.
There were also account-level side effects afterward, including Actions/Pages being broken until the mess got sorted out.
For a while several FOSS repos I maintain were just gone from the public internet because of this.
That's the part I care about most.
These tools are starting to get enough access that when they screw up, they can screw up outside the thing you were actually asking them to touch.
I was trying to work on code.
Instead I got a GitHub suspension, dead OAuth tokens, broken Pages, and a support archaeology project.
OpenAI support, weirdly, was the least shitty part of the whole thing.