*OpenAI and Google face new questions about their AI image generators after researchers successfully tricked several models into creating nude images that their safety systems should have blocked.
The findings highlight a potential risk for people whose photographs could be misused to create sexualized deepfakes without their consent. However, the study did not document actual victims.
According to Startup Fortune, Cybersecurity researchers at White Circle examined whether the companies’ image tools could resist carefully designed attempts to override their restrictions. Their results revealed an unexpected problem: Some recently released models performed worse than their predecessors.
Google’s Nano Banana 2.1 failed 20.7% of the researchers’ attack attempts, compared with 9.5% for Nano Banana 2.
Scroll to continue reading
OpenAI’s GPT-Image-2.5-Flare recorded a 7.8% success rate, up from 3.3% for GPT-Image-2. An older OpenAI model, DALL·E 3, registered 13.3%.
Those percentages reflect the researchers’ specialized testing, not the likelihood that everyday users will encounter explicit images.
The team developed an automated technique called RISE, which explores different ways to get around image-generation restrictions. Unlike approaches that repeatedly adjust individual requests, RISE searches for broader strategies that can produce multiple attempts.
Researchers then evaluated the generated images using criteria calibrated against human reviewers. An attack counted as successful if at least one of three images violated the nudity rules. The findings follow growing scrutiny of AI-generated sexual content and its potential misuse.
White Circle’s published study does not determine why the newer models were more susceptible. Changes in image-generation technology or content moderation could help explain the differences.
The research also raises a question for companies developing these systems: How can they identify vulnerabilities before people exploit them?
For now, the results demonstrate that even widely used AI image tools can produce prohibited material when their safeguards face targeted attacks.
MORE NEWS ON EURWEB.COM: Cardi B Goes Viral Cleaning Her Grandparents’ House, and Fans Have Jokes
Sign up for our Free daily newsletter HERE
