A group of cybersecurity researchers has discovered a prompt that slips past ChatGPT’s guardrails, forcing the artificial intelligence to generate disturbing images.
The finding, revealed in a BBC podcast episode, raises urgent questions about the limits of safety training in large language models and how those vulnerabilities could be exploited.
“Cybersecurity researchers discovered a prompt that bypasses ChatGPT's guardrails to create disturbing images.”
ChatGPT, built by OpenAI, is designed to refuse harmful requests — but the researchers found a way to trick it. The BBC’s Chris Vallance, who presented the episode, said the prompt “gets past ChatGPT’s guardrails and causes it to generate some disturbing images.” The exact nature of those images and the wording of the prompt have not been made public, but the discovery underscores the cat-and-mouse game between model developers and those probing their defences.
The episode, broadcast on 16 June 2026 as part of the BBC World Service’s Tech Life programme, unpacked what the finding tells us about the way AI is trained and how it could be exploited. The researchers, whose affiliation was not named, belong to the cybersecurity community that routinely tests the boundaries of commercial AI systems.
This is not the first time ChatGPT’s safety measures have been circumvented. Previous research has shown that so-called “jailbreak” prompts can make the model ignore its training to give dangerous advice or produce offensive content. But the latest discovery is notable because it targets image generation — a newer capability of ChatGPT that has been heavily restricted following public outcry over earlier incidents of deepfakes and violent imagery.
The BBC episode also touched on other technology stories, including a conversation with the UK’s ministry of transport chief scientific adviser about potholes and a company that has put a quantum diamond magnetometer into space to measure magnetic north. But the ChatGPT vulnerability dominated the discussion, with Vallance and producer Imran Rahman-Jones framing it as a warning that might influence how AI companies approach safety testing in future.
The question now is whether OpenAI can patch the loophole before it is used more widely — or whether the very architecture of large language models makes such exploits inevitable.