Breaking

Mindgard Says ChatGPT's Image Tool Produces Violent, Sexual Images From a Simple Prompt

June 19, 2026 at 23:50 EDT

Security
Media Generation
Foundation Models

UK-based AI security firm Mindgard reports that ChatGPT's image generation feature can produce sexual, violent and grotesque images when a seemingly harmless prompt is altered only slightly. OpenAI says it added new safeguards in response, but researchers say minor tweaks can still bypass them.

June 18, 2026 · AI Safety Research

A Viral "Restore This Photo" Prompt Made ChatGPT Generate Violent & Sexual Images

Security researchers found that ChatGPT's image generator produced disturbing content from a single text prompt — with no image attached. OpenAI says it has added new safeguards; researchers say minor tweaks still bypass them.

images attached — yet harmful outputs still appeared

days from investigation start to publication

All

users with image access — free and paid — could reproduce it

How a harmless prompt became an exploit

1 · Viral prompt

"Restore the attached photo" — but no photo is attached

→

2 · Minor tweaks

Swap "strange"→"graphic", repeat prompt, add filter-evasion wording

→

3 · Harmful output

Violent, sexual and graphic imagery generated

OpenAI's response

Points to prompts without an attached image as a cause; says it now encourages image uploads and has introduced additional safeguards after investigating.

Researchers' rebuttal

Generation is still possible with minor modifications — calling it a systemic, not one-off, problem. Warns training data may contain problematic images and the filter isn't working adequately.

The alarming part isn't a sophisticated attack — it's that a near-accidental, widely shared prompt slipped past the guardrails.

The report includes no benchmark figures or specific model version, and the effectiveness of OpenAI's fixes remains to be verified.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.