Hey there
Thank you for taking the time to share such a detailed report. I can hear how frustrating this has been, especially when you’re simply trying to run very basic prompts like “Hello” and are getting blocked by the system. I’m really sorry you’ve had to deal with repeated invalid prompt errors and responses that didn’t feel useful or relevant.
To provide some clarity: prompts are automatically scanned against a set of classifiers that look for patterns associated with misuse (for example, jailbreak-like structures, distillation attempts, or certain bio/chemistry workflows). These safeguards are important, but sometimes they sweep up perfectly safe prompts by mistake — which is why something as simple as a “Hello” can trigger an error. Enforcement can also differ by model: stricter checks are applied to reasoning variants (like o3, o4-mini) compared to others, so the same input may work in one model and fail in another.
The challenge is that the system currently reuses the same “invalid prompt” error message whether the block comes from:
content-level safety rules,
classifier triggers (e.g. distillation/jailbreak detection), or
account-level restrictions (for example, if streaming is attempted while an account is flagged as WARN).
That’s why the message can feel vague and unhelpful it doesn’t explain which mechanism was responsible in your specific case.
We deeply appreciate your patience here, and I'm going to flag this to our team internally, unfortunately I don't have a timeline for when this issue will be resolved.