OpenAI and Google image models easily tricked into making nudes
A White Circle preprint says the safety filters of DALL·E 3, Nano Banana 2 and GPT-Image-2 can be bypassed with off-the-shelf tools.
Policy#02
13%
Best-case bypass rate on DALL·E 3
Image: Ars Technica
13%Best-case bypass rate on DALL·E 3
The short version
- White Circle researchers reported flaws in the guardrails of DALL·E 3, Google's Nano Banana 2, and GPT-Image-2 in a preprint.
- In the best case, about 13 percent of the attack prompts on DALL·E 3 produced images that human reviewers confirmed as explicit.
- The rate was under 10 percent for Nano Banana 2 and about 3 percent for GPT-Image-2.
- White Circle told Ars that the newer ChatGPT Images 2.5 and Nano Banana 2.1 are significantly more breakable than their previous versions.
Why it matters
The findings suggest that safety defenses on leading image generators may lag behind their capabilities. The researchers also said the method did not need custom-built or frontier models, which lowers the barrier for attackers.
“Critically, none of these requests required custom-built or frontier AI models,”