Researchers report gaps in AI image-generation safeguards

Ars Technica reports that White Circle researchers bypassed explicit-image safeguards in tested OpenAI and Google models. Human reviewers found explicit outputs in about 13% of attacks on DALL·E 3, under 10% on Nano Banana 2 and about 3% on GPT-Image-2. These are attack-test rates, not real-world prevalence; the companies did not answer Ars’ requests for comment.