Researchers: New OpenAI and Google Models Easily Tricked Into Nudes

Red-teaming conducted in recent months suggests that the newest AI image generators from Google and OpenAI can be tricked into producing non-consensual sexual imagery — and that as these models grow more capable, they may be getting worse, not better, at blocking explicit outputs.

That is the warning from safety researchers, who say a large volume of AI-generated explicit content is now spreading unchecked while regulators do little to meaningfully stop it. The findings land on the two companies whose image models sit at the center of the consumer AI market, raising the stakes for both platform policy and enforcement.

What did the red-teaming find?

According to researchers cited in the report, recent adversarial testing of leading image generators from Google and OpenAI indicates that the newest, most capable model versions are more readily manipulated into producing nude imagery than their behavior policies would suggest. The testing was carried out over recent months.

The core claim is counterintuitive. Capability and safety are not moving together. As the image generators improve at following user instructions, they appear to become easier to steer past the guardrails designed to block explicit content.

The report attributes the finding to safety researchers who continue to warn about the scale of the problem. Their conclusion is direct: the volume of AI-generated explicit imagery is large, and it is spreading without meaningful checks from either the companies or regulators.

Where does the non-consensual content end up?

A Google co-authored study published this July in Computers in Human Behavior surveyed internet users across the United States, Australia, and the United Kingdom. It found that AI-generated non-consensual sexual imagery is commonly shared in text messages and direct messages, or stored in private clouds.

Those channels matter for a practical reason. Content in DMs and private storage is far less likely to be detected or reported than material posted on open, searchable platforms. The imagery circulates outside the view of both content moderators and researchers who rely on public-facing data.

The study's authors — Google researchers among them — reached a clear conclusion about where to intervene. The "most critical intervention point" remains at the source: the harmful AI tools used to make the fake sexual images.

That framing puts the responsibility on model providers such as Google and OpenAI rather than on downstream platforms. If the tools can be tricked into producing the imagery in the first place, policing its distribution after the fact is a losing proposition — especially when distribution happens in channels that are invisible by design.

Why does this matter now?

Two forces make the timing significant.

First, capability. Google and OpenAI are shipping their most advanced image generators to date, and the red-teaming results suggest the safety layers around those systems have not kept pace with their generative power. A model that follows instructions more faithfully also follows malicious instructions more faithfully.

Second, enforcement. The researchers' warning is explicit: regulators are doing little to meaningfully stop the spread of AI-generated explicit content. No binding intervention currently forces providers to close the gaps that red-teamers are documenting.

The combination — improving models, weak blocking, minimal regulation, and distribution channels that resist detection — defines the problem the researchers describe.

What is at stake for Google and OpenAI?

Both companies publish usage policies that prohibit non-consensual sexual imagery. The red-teaming findings suggest a gap between policy and model behavior at the very frontier of their image-generation product lines.

The July study's conclusion sharpens that pressure. If the "most critical intervention point" is the tool itself, then every image generator that can be tricked into producing nudes is the point of failure — not the messaging apps, cloud storage services, or forums where the output later circulates.

For companies that co-author research on the harms of these tools while simultaneously shipping the models in question, the tension is now documented in their own findings and in independent red-teaming alike.

The open question is whether the next round of model releases closes the gap. The researchers' evidence suggests the trend so far runs the other way: more capable generators, easier tricks, and a regulatory response that has not yet materialized.