Cyber Abuse deactivation appears to be a classifier false positive - requesting human review (Case ID: C-zjz01nwUQe05)

Hi OpenAI team,

Reporting what appears to be an automated classifier false positive. My ChatGPT account was deactivated for “Cyber Abuse,” and my appeal was rejected with a templated message (“we are upholding our decision… We
will no longer consider additional requests to appeal this case”). The rejection arrived as a form reply — I’ve seen the identical sentence quoted by other users here for unrelated cases — so I don’t believe a
human reviewed it.

Case ID: C-zjz01nwUQe05

Root cause of the false positive: I work in cybersecurity recruiting. I was pasting candidates’ resumes into ChatGPT to screen them for open red-team/offensive-security job requisitions. Those resumes contain
terms like “penetration testing,” “exploit development,” and “attack automation” because they describe the candidates’ own prior, client-authorized work — not any request from me to perform such activity. No live
system was ever in scope; the only content processed was candidates’ employment-history text.

Per OpenAI’s own Cyber Abuse Policy, prohibited conduct is activity that “enables real-world harm, targets live systems without authorization, or provides actionable assistance for abuse.” None of these occurred
— the model was only asked to summarize/classify text the candidates themselves wrote. All hiring decisions were made by a human recruiter; ChatGPT was only a triage aid.

I’m also aware of a broader wave of “Cyber Abuse” false-positive deactivations reported here recently, which points to a classifier calibration issue.

Could a staff member please escalate this for a human review? I’m happy to provide job postings, ATS records, account domain ownership, and the actual prompts used.

Thank you.

Thanks for flagging this issue @tanwe

I’ve sent you a message on support ticket 11028403. To avoid duplication and keep the investigation in one place, I’ll close this thread. All further communication will take place through the support ticket.

~ Smith