What changed
OpenAI released `omni-moderation-csam-latest` through the Moderation API for approved organizations at no charge. It returns separate risk signals for potential child sexual abuse material and potentially AI-generated CSAM so trained reviewers can prioritize images that are not already known by hash.
Why it matters
Novel and synthetic abusive imagery can evade known-content hashes, but automated scores also carry severe legal and human consequences. The useful change is a restricted triage signal with a model card, not an automated verdict.
What to do
Eligible trust-and-safety, child-protection or law-enforcement teams can apply for access, validate thresholds on a legally governed, representative holdout, and keep hash matching, trained human review, evidence handling and reporting decisions separate from the classifier score.
Caveat
Access requires application, verification, onboarding and terms. OpenAI says the score does not confirm illegality, provenance or a reporting obligation; video is untested, some subcategory signals remain internal, and independent benchmarking is still needed.