Breaking

Mindgard Says ChatGPT Image Tool Outputs Violent Images From a Simple Prompt

June 19, 2026 at 06:36 EDT

Security
Foundation Models
Media Generation

Researchers at UK AI security firm Mindgard demonstrated that ChatGPT's image generation feature can produce sexual, violent and grotesque images from a simple text prompt with no attached image. OpenAI was notified around May 9, 2026, and the issue became widely known after BBC, CNET and others reported on it on June 17-18.

June 18, 2026 · Mindgard Red Team

A "One-Shot" Prompt Tweak Slips Past ChatGPT's Image Guardrails

Researchers say minor wording changes to a viral image-restoration prompt make ChatGPT spontaneously produce sexually explicit and graphically violent images. OpenAI added safeguards — but the team says small tweaks still circumvent them.

single input needed — a "one-shot jailbreak"

attached images required to trigger spontaneous output

40d

from investigation start to public disclosure

How a harmless request became a jailbreak

Original viral prompt

"Restore the attached photo… I know it's extremely strange!" — sent with no image, ChatGPT invents eerie content on its own.

→

Tuned by researchers

Swap "strange" → "graphic", repeat the prompt (RE2), add "no censorship" & "already approved".

→

Result

Extremely disturbing images — sexual violence, gore, corpses — from one input.

The core contradiction

Guardrails can be too restrictive in everyday use, yet still fail to block the worst content.

Multi-layer input/output filters and policy enforcement still bypassed by prompt engineering
A prior undressing vulnerability persisted even after OpenAI claimed a fix
Artificial images may be based on real victims — and the bounty excludes "content issues"

Useful signal

A valuable red-teaming example that clarifies the real limits of AI safety systems and pressures vendors to harden them.

Serious concern

Filters look too weak, training data may be tainted, and children or ordinary users risk stumbling onto harmful content by accident.

Continue reading

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