Researchers at British AI security startup Mindgard have shown that the latest public version of ChatGPT can be made to generate sexualised images and scenes of graphic violence through slightly altered prompts.
The firm identified the issue by modifying a widely shared instruction originally intended for humorous results.
The vulnerability affects OpenAI’s GPT-5.4 model. The BBC reported that, after being contacted, OpenAI said it had investigated the trend and introduced additional safeguards against such prompts.
The company stated that it maintains multiple layers of protection to prevent content that breaches its terms and conditions.
Mindgard researchers reported that further small changes still allowed the generation of concerning content.
Founder Peter Garraghan, a professor in the computing department at Lancaster University, described some outputs as very gruesome and sometimes sexualised.
Read: ChatGPT 1 Billion Users Milestone Sets App Record
Jim Nightingale, Mindgard’s AI safety and security researcher, said he was shaken by the images produced. OpenAI reiterated that its policies prohibit sexual violence, non-consensual intimate content and extreme gore.
The company combines automated systems and human review to block harmful material. Researchers first alerted OpenAI in May.
