So basically, table stakes for running a user-submitted-content site will be to pay for your own GPU node/cluster and manage your own inference stack (running a worse model than you could use in the cloud). Otherwise you're at the mercy of opaque and kafkaesque (by necessity) service termination without warning for CSAM (real or hallucinated), and subsequent refusal to communicate about anything related to it.
[1] either by queuing all posts for classification before showing them publically, or just best-effort after-the-fact detection of spam or abuse, and upon detection they're moved to a spam queue.
Or just use one of the many third party inference providers that markets an image classification endpoint intended for user generated content.
The fact that people have autonomous agents making these kinds of decisions is still mind-boggling to me. Maybe I'm stuck in the past. But could this agent have been racking up things of dollars in charges because of this decision GP didn't even know about?