A group of cybersecurity researchers has discovered a prompt that slips past ChatGPT’s guardrails, forcing the AI to generate disturbing images. The finding, revealed in a recent episode of the BBC’s Tech Life programme, raises fresh questions about the limits of safety measures built into large language models. Presenter Chris Vallance unpacked what the vulnerability tells us about the way AI is trained and how it could be exploited by bad actors.
The researchers, whose work was discussed on the show, did not disclose the specific prompt publicly, but the episode’s summary notes that the exploit causes “some disturbing images” to be produced. The discovery comes as chatbots like ChatGPT face increasing scrutiny over their ability to refuse harmful requests. The programme suggested that the incident highlights broader weaknesses in how AI systems learn to block inappropriate content – and how determined users can find ways around those blocks.
“Cybersecurity researchers found a prompt that bypasses ChatGPT's guardrails to generate disturbing images.”
Elsewhere in the episode, the team turned to a more prosaic problem: potholes. After a previous Tech Life episode on the subject, the UK’s ministry of transport contacted the show. Its chief scientific adviser joined Vallance to discuss potholes and the future of transport, offering a government perspective on the nation’s crumbling roads and potential technological fixes.
The show also ventured into cutting-edge physics with a segment on quantum diamond magnetometers. A company that has just launched one of these devices into space explained how it works – measuring the precise location of magnetic north. The technology, the guest said, could transform navigation systems that rely on magnetic field data.
The episode left listeners with an unsettling takeaway: while AI can be coaxed into breaking its own rules, the same determination that drives cybersecurity researchers might also be used by those with malicious intent. How long before the next prompt slips through?