OpenAI reveals tactics to fight child sexual exploitation

As AI models get better at seeing images and spinning complex stories, bad actors are finding increasingly disturbing ways to test their boundaries.

OpenAI just pulled back the curtain on how people are attempting to abuse ChatGPT for child sexual abuse material (CSAM), detailing the exact evasion tactics they are catching and the defensive stack they use to block them.

Novel abuse tactics beyond image generation

Most people assume preventing child exploitation in AI is just about blocking the image generator from producing illegal photos. The real battlefield, according to OpenAI, has moved far beyond that.

One tactic OpenAI highlighted involves users uploading illicit imagery directly into ChatGPT, asking the model to write meticulous descriptions of what is depicted. In other cases, users attempt to coax the model into sexual roleplay narratives while uploading illegal material, or try to get ChatGPT to write fictional stories involving the abuse of minors.

To counter this, OpenAI is using hash matching against vetted databases from Thorn, running Thorn’s specialized classifiers on uploaded content, and pairing them with internal context-aware AI monitors. Anyone attempting these prompts is immediately banned, with incidents referred to the National Center for Missing and Exploited Children (NCMEC).

The crackdown also applies directly to developers. If a third-party app targeting minors fails to ban problematic users after a warning, OpenAI shuts down the developer account entirely.

The legal paradox of red teaming

OpenAI also pointed out a massive hurdle holding back its safety work: it is technically illegal for safety teams to stress-test their own models against CSAM.

Because possessing or creating this material is strictly illegal under U.S. law, AI researchers cannot create or use even simulated test data for red teaming. That means safety teams are forced to build defenses blindfolded, without the rigorous vulnerability probing standard for other cybersecurity threats.

That legal catch-22 is why OpenAI is pushing for legislative updates, backing bills like New York’s Child Sexual Abuse Material Prevention Act to establish legal protections for proactive threat detection and research cooperation.

Why it matters

Multimodal AI turns input comprehension and narrative writing into potent tools, and bad actors are treating vision-enabled chatbots as automated processors for illegal material.

Publishing these attack patterns is a smart move. It forces the broader AI ecosystem to look beyond simple keyword blocks and recognize that text-only fiction and image-to-text analysis are active attack surfaces.

More importantly, it exposes how outdated safety laws currently hamstring defense. When security researchers risk federal crimes simply by testing whether their safety filters work, the bad guys hold a structural advantage. For developers building on top of frontier APIs, the message is loud and clear: platform immunity is gone, and child safety compliance must be baked directly into your application layer.