Article 272: Off-the-Shelf Jailbreaks Are Beating Image Guardrails at OpenAI and Google

Researchers at White Circle — a public-benefit outfit staffed by alumni of Anthropic, DeepMind, Hugging Face and OpenAI — say the leading text-to-image models are getting worse at blocking explicit output, just as they get more capable. Using only publicly downloadable tools and standard API calls, their attack produced human-confirmed explicit images at rates around 13 percent on DALL·E 3, under 10 percent on Google's Nano Banana 2, and roughly 3 percent on GPT-Image-2. Crucially, none of the prompts required a jailbreak in the traditional sense.

Key Takeaway: A 3 percent success rate sounds like a rounding error until you remember it is measured per attempt and the cost of an attempt is near zero. Safety that holds at 97 percent against a scripted, off-the-shelf attack is not a guardrail — it is a tax on the adversary, and a cheap one.

Source: Ars Technica, 14 October 2026