OpenAI adds safeguards after graphic image loophole found

Researchers have discovered that OpenAI's latest public version of ChatGPT can be manipulated to generate graphic and sexualized images using seemingly innocuous prompts, prompting the company to implement additional safeguards. The findings, reported by British AI security startup Mindgard, highlight persistent challenges in securing advanced AI models against misuse.

Mindgard's researchers found that by slightly altering a widely shared instruction originally intended for humorous results, they could prompt ChatGPT's GPT-5.4 model to produce disturbing content without directly requesting it. The resulting images included depictions of severe head injuries, a young woman covered in blood with features suggesting sexual violence, and a frightened woman tied up and gagged in a dirty room. The chatbot labeled some outputs with titles like "Grim crime scene aftermath" and "abandoned in fear and restraint." Other generated images showed nudity and sexual posing.

Mindgard founder Peter Garraghan, also a professor at Lancaster University, described the images as very gruesome and sometimes sexualized. He expressed concern that the prompt did not specify such subject matter, yet the AI...