Researchers find gaps in OpenAI and Google image safeguards
The tests show gaps in safeguards on widely available image tools, which a Google co-authored study identified as a critical point for limiting nonconsensual sexual imagery.
• White Circle says its prompt attack produced explicit images from OpenAI and Google models.
• In its tests, human reviewers confirmed explicit outputs from about 13% of DALL·E 3 attack prompts; rates were lower for two other models.
• White Circle says ChatGPT Images 2.5 and Nano Banana 2.1 were more susceptible than their predecessors under the same method.
Read more
October 14, 2026 25
