Safety researchers warn that AI-generated explicit content is spreading largely unchecked, while regulators do little to stop it. A Google co-authored study of internet users in the United States, Australia, and the United Kingdom found that such non-consensual content is often shared through private messages or stored in private clouds, where it is less likely to be detected or reported. The researchers concluded that the most critical intervention point remains the harmful AI tools used to create fake sexual images.
Red-teaming in recent months suggests that leading image generators from Google and OpenAI may be getting worse at blocking explicit outputs as they become more capable. White Circle, a public benefit corporation backed by senior leaders and researchers from Anthropic, DeepMind, Hugging Face, and OpenAI, has raised the alarm. In a preprint published last month, its researchers exposed flaws in the guardrails of DALL·E 3, Google's Nano Banana 2, and GPT-Image-2. Using off-the-shelf tools and API calls, they bypassed safety filters and generated explicit content about 13 percent of the time, a bypass rate they describe as the highest of any method tested under the same standard.
White Circle also reports that newer models, ChatGPT Images 2.5 and Nano Banana 2.1, are "significantly more breakable" than their predecessors. The company plans to open-source its attack pipeline and present the research at the NeurIPS 2026 conference in November. Neither Google nor OpenAI responded to Ars Technica's requests for comment.
CEO Denis Shilov said the findings are concerning because the same weakness could affect defenses being updated to block other harmful AI behavior, such as hacking. The company's key message is that AI safety defenses at leading labs are weaker than they acknowledge.
Source: Ars Technica · Summarized by HeadlinesBriefing