OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say

Safety researchers say a lot of AI-made explicit images is still spreading with very little meaningful pushback from regulators. In many cases, the stuff is passed around in direct messages or tucked away in private clouds, which makes it harder to spot and report. That was one of the findings in a Google co-authored study of internet users in the United States, Australia, and the United Kingdom published in July.

The bigger point from that work is blunt: the best place to stop this kind of abuse is at the source. If the tools used to generate the fake sexual images are weakly protected, everything downstream gets uglier fast. Google’s own researchers called that the “most critical intervention point.”

But recent red-teaming suggests the companies making some of the most capable image generators may be losing ground on that front. Tools from Google and OpenAI are becoming more impressive in what they can create, and at the same time, researchers say they may be getting easier to steer into explicit output.

That’s the annoying shape of this problem. The models get better, the guardrails get shakier, and the harm stays cheap to produce and easy to move around in private channels. The result is a nasty little race where the people building the systems seem to be chasing their own failures.