GPT 6 thinks all my prompts are violating OpenAI’s usage policy.
This is very strange and shouldn’t happen, I can’t use GPT 6 at all right now.
Does anyone know how to fix this?
Oh wow, have you tried logging out and back in again? (more likely a server issue you can do nothing about though!)
Very strange!
I’ve tried many times. It doesn’t work.
Any prompts before the ones in the screenshots?
Any emails from OpenAI? Might want to hit up support to let them know.
Free version or Pro? I’d definitely reach out to support.
Welcome to the developer community, @AlopexG.
That behavior does not appear to be expected.
To help narrow down the cause, could you confirm whether the affected chats are associated with a project and whether Memory is enabled under Settings → Personalization?
Hello @sps, thanks for the reply. It is not in a project, memory is not enabled, and there are also no custom instructions,
Thank you for confirming those details. I’d recommend opening a support ticket at help.openai.com and sharing the case number here so the team can investigate further.
Hey all,
Since around midday today (~12:56 my time) Codex has been rejecting every single message I send. Doesn’t matter what it is — a code review, “fix this typo”, even just “hello” — I get this back every time:
Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different prompt: Reasoning models | OpenAI API
The underlying error is a 400 with code “invalid_prompt” (type invalid_request_error). Model is astral.
Things I’ve already ruled out:
- It’s not my machine or my config. I tried from a completely different device with a fresh Codex setup, different settings, and got the exact same thing. So whatever this is, it’s tied to my account (or account + model), not anything local.
- It’s not the content. The prompts are as boring as it gets. Nothing remotely close to policy territory.
- Not rate limited or logged out — the account itself is fine, I can log in, no emails from OpenAI, nothing in the notifications.
I’ve seen a few older threads about this same error popping up on reasoning models over the past year or so (o1 era, then again on GPT-5), and it looked like a moderation false positive on the server side each time. Feels like the same thing happening again with astral, but I honestly can’t tell if it’s a broad outage or just my account got flagged somehow.
Couple of questions:
- Is anyone else seeing this on astral right now, or is it just me?
- If you hit this before — did it clear on its own, or did you have to go through support? How long did it take?
- Should I be worried about this turning into an actual ban? I’ve read a couple of horror stories where people kept retrying and then got their access pulled. I’ve stopped sending anything to astral for now just in case.
Screenshot of the error attached.
Thanks.
I have same problem
only astar
Invalid prompt: your prompt was flagged as potentially violating our usage policy. Please try again with a different prompt: Reasoning models | OpenAI API
I created an account specifically to report this bug as well. Whenever I send any prompt using GPT-6, it immediately triggers a safety warning and stops the task. The only way it works properly is if I downgrade to GPT-5.6. GPT-6 is completely unusable right now—even a simple greeting like “hello” fails.
For those reporting the problem also consider using Codex feedback.
same problem, why hasn’t anyone resolved it
5.6 had the same problems on release and it took openai a few weeks to ease up on the ridiculous guard rails it seems with a new model release openai over stacks the model with guard rails
Thanks - I followed the /feedback procedure and uploaded a minimal diagnostic session for what appears to be the same issue.
GPT-6 Astra worked earlier today on my ChatGPT Pro account, then started rejecting every prompt with invalid_prompt, including a clean isolated prompt equivalent to “reply only OK.” The same isolated test with GPT-5.6 Sol returns OK.
Environment: macOS, bundled Codex CLI 0.153.4. The failure remained after excluding user config, project rules, plugins, and repository context. Usage was 0%.
Uploaded feedback/session ID: 01a07fcf-a4b1-7112-8e2e-6ebb1213fe13
I also attached the same diagnostic ID to openai/codex issue #43237.
Since multiple users report the same Astra-only behavior across machines and accounts, this appears consistent with a server-side routing or safety-check regression. Any escalation to the Codex/Astra team would be appreciated.
Welcome to the forum!
FYI
I do not think you can do any better than you have done using both /feedback and replying in an open issue with details.
No one at openai seems to be working on this, including the forums and support staff，Frustrating, $200/month subscription without access to the latest features
I’m wondering if something else is being included in the context sent to the model—perhaps something from memory—and that is now triggering the response.
OpenAI’s Unrolling the Codex agent loop explains that the user’s input is included in a larger prompt prepared for the model. That prompt can contain other instructions and context in addition to the text the user entered.
So even though the visible prompt may be only “hello,” that does not necessarily mean “hello” is the only information the model is receiving.