Report says AI systems can be tricked into extremist bomb-making guidance

A report released this month says artificial intelligence systems can be manipulated into providing coaching on bomb-making and weapons, raising fresh concerns about extremist misuse. The findings come as a Queensland teenager faces extremism charges linked to the alleged use of AI to create mass-shooting scenarios. The report describes the issue as a major national security concern and argues that conversational AI can be used in ways that go beyond ordinary web searches.

Sponsored

The United Nations-supported group Tech Against Terrorism tested 27 AI models with prompts that could be used for terrorism, including step-by-step requests for explosive devices and 3D-printed firearms. According to the report, ChatGPT refused only 48% of more than 2,000 harmful requests entered by testers. The group's executive director, Adam Hadley, said the risk was heightened because AI systems are designed to engage in back-and-forth conversation and refine answers over time.

The report's release has added to scrutiny of how AI platforms respond to extremist or violent prompts. A spokesperson for the company behind ChatGPT said it worked continuously to make the platform safer by stopping it from providing harmful information. Hadley said models that iterate with users and answer follow-up questions behave less like a manual and more like a coach, which he said makes them more dangerous in the hands of extremists.

The Australian case cited alongside the report has become part of that wider concern. In Queensland, a 13-year-old boy is alleged to have used AI on multiple occasions before his arrest to simulate a mass shooting. A court heard that the teenager generated a note titled "The Albert Massacre", apparently referring to Albert State School in Maryborough.

Sponsored

Police also allege he asked the program to create a mass-shooting story in a style linked to a Bondi Beach shooting and a Jewish and Black festival. The report also points to a broader pattern of concern about AI and radicalisation. It says violent extremists can exploit systems that are built to respond conversationally, rather than simply returning static search results.

Hadley said the use of AI by extremists is ubiquitous, and he also linked the issue to a stabbing attack in a Melbourne shopping centre that is alleged to have involved AI radicalisation. What remains unclear is how widely these weaknesses affect different models and what safeguards are most effective in practice. The report calls for stronger action, including geoblocking systems that breach Australian safety laws, but it does not set out a final regulatory solution.

The immediate question for authorities and developers is whether current safety controls can prevent AI tools from being used to generate extremist or violent material at scale.

#Report #Says #Systems #Can #Tricked

Sponsored