TL;DR — An API key of ours was used by a third party for 181,618 requests across 36 models on 2026-08-31 and 09-01, costing USD 2,825.61 on a project that had spent USD 61.28 in the previous thirty days. I reported it twice through the “Report unauthorized activity” branch of openai.com/form/appeal/, the second time because Support told me to. Both times I received an identical automated email titled “Appeal Decision” saying a prior decision “stands” — for an appeal I never filed, on an account with no prior decision. Support then told me it cannot confirm whether a review happened, cannot issue a reference number, and cannot confirm the case is open — but that if a review did happen, the outcome is final and unappealable.
I’m looking for the route where a case like this actually gets seen.
What happened
The key served one internal service, a self-hosted LightRAG instance configured to call only gpt-4.1-mini and text-embedding-3-small. Usage straight from the Usage export:
Date Requests Distinct models
2026-07-24 69 1
2026-08-27 70 1
2026-08-28 4,307 7
2026-08-29 2,598 7
2026-08-30 143 7
2026-08-31 45,628 32
2026-09-01 135,990 36
Short version of why it wasn’t us: the key never touched more than 7 models on a normal day and reached 36 under the abuse, including gpt-image-2, gpt-4o-mini-tts and o3-pro, which that service cannot call at all. Our host’s logs show it received zero HTTP requests during the window, and its knowledge graph is unchanged either side of it. Consumption stopped when the key was deleted and has not resumed. Nothing else changed.
The Logs view is what convinced me this was resale rather than a runaway job of ours. Filtered to this key over 08-27 to 09-01 it returns 14,277 entries, full of Reply with exactly: OK probes and a Chinese-language liveness check — the shape of a pool verifying a stolen key still works — alongside unrelated third-party workloads. One entry is billed to o4m-sonic-p-cc-api-ev3, a model string I cannot find in the public model list.
Happy to share the full per-day, per-key and per-project exports with staff.
The part I’d like help with
2026-09-01 — submitted through openai.com/form/appeal/, choosing “Report unauthorized activity involving my account” › “My API key was compromised or used without permission”.
Twelve minutes later — email from noreply@tm.openai.com, subject “OpenAI - Appeal Decision”:
Thank you for reaching out. We have already reviewed your appeal and the original decision stands.
I never filed an appeal. There is no prior decision on this account — no warning, no deactivation, nothing to uphold.
Escalated through help.openai.com. I attached the full evidence and pointed out that the form I was being sent to was the same one that had already auto-replied, and that it was OpenAI’s own AI-assisted support that sent me there originally. The reply was a template telling me to submit through that same form.
2026-09-06 — resubmitted exactly as instructed. The description opened with “RESUBMISSION REQUESTED BY OPENAI SUPPORT — please do not close as duplicate”.
The same day, the identical email arrived again. Word for word, same subject, same sender. The ticket was then closed, despite my having asked in writing that it stay open until the resubmission was confirmed.
A support specialist later confirmed the submissions had been received — the first confirmation in two weeks — but then told me:
I don’t have access to confirm the internal review status, provide a case or reference number, confirm whether the case remains open, or provide an estimated review timeframe.
and, in the same message:
If the review has already concluded, the communicated decision is considered final, and there isn’t a manual appeal or override process available through Support.
I don’t think the agent did anything wrong; this reads like the limit of what Support is able to do. But the resulting position is that a decision which cannot be confirmed to exist, whose content was never communicated to me, and which carries no reference number I can cite, is nevertheless final and not appealable.
The content-report form doesn’t cover the API
I also checked whether openai.com/form/report-content/ was the right destination instead, since a third party using a stolen key looks more like platform abuse than a billing dispute. It isn’t: its required “Which product or feature does this relate to?” field lists ChatGPT, GPTs/GPT Store, Operator, Codex, Ads and similar — there is no API option. It’s also built around content at a URL, and API traffic has none.
Questions
- Is
/form/appeal/actually the destination for a compromised-key case, or does it only serve enforcement appeals despite offering the unauthorized-activity branch? - Has anyone here received something other than the appeal template from that branch?
- Is there any way to obtain a reference number for a submission made through it? Nothing I’ve received in two weeks contains one, and Support says it cannot issue one — so every contact restarts from zero.
- If Support cannot confirm that a review occurred, on what basis is its outcome final?
- Where does a compromised-key case actually get human review?
I’d genuinely like to be wrong about the routing. If the path exists and I simply haven’t found it, please point me at it.