OpenAI and Google AI models vulnerable to explicit content prompts

Research indicates that advanced AI image models from OpenAI and Google remain vulnerable to adversarial prompts designed to bypass safety guardrails. A test by AI safety firm White Circle found that OpenAI’s GPT-Image-2.5 Flare and Google’s Nano Banana 2.1 were more susceptible to prompts seeking female nudity than their predecessors. Specifically, GPT-Image-2.5 Flare had a prompt-level success rate of 7.8%, while Nano Banana 2.1 reached 20.7%.

Other studies have highlighted various methods for bypassing restrictions. Research from Mindgard demonstrated that minor adjustments to prompts could trigger graphic imagery in OpenAI models, while Icaro Lab found that “adversarial poetry” could successfully elicit forbidden content from a majority of tested chatbots.

The social impact of these vulnerabilities is significant. A study involving users in the US, Australia, and the UK reported that 6.9% of respondents experienced image-based sexual abuse specifically tied to AI-generated content. Furthermore, a Common Sense Media survey found that nearly half of US teens had encountered AI-generated sexual material, with 24% stating the content depicted themselves or people they knew.