Mindgard finds ChatGPT GPT-5.4 tricked into creating graphic sexualized images
A U.K.-based AI security firm, Mindgard, found that OpenAI's ChatGPT can be tricked into making graphic and even sexualized images just by tweaking popular prompts.
Their tests showed the GPT-5.4 model could create violent and morally troubling content, raising big questions about how easily AI tools might be misused.
Mindgard shared disturbing images with BBC
Mindgard shared some of the images with the BBC, including a badly injured person and a crime scene suggesting sexual violence.
The founder, Peter Garraghan, said he was worried by what the AI could do on its own; Jim Nightingale, the company's AI security and defense researcher, admitted he felt shaken after seeing these pictures.
OpenAI pledges safeguards, Mindgard found workarounds
OpenAI says it's working on tougher safeguards using layered systems to block this kind of content and monitor misuse.
Mindgard researchers already found ways around them.
Even as OpenAI promises to keep improving its security measures, experts warn there are still serious ethical risks in how AI might be used.