Artificial intelligence is vulnerable to being exploited by violent extremists, according to a report that flags the threat as “a major national security concern”.
Last month a young teenager was charged in Queensland over allegations he created alarming scenarios on an AI platform.
A court heard the teenager generated a note titled “The Albert Massacre”, seemingly in reference to Albert State School in Maryborough. (Supplied)
The United Nations-supported not-for-profit Tech Against Terrorism asked 27 AI models for information that could be used for an act of terrorism, including step-by-step instructions for making explosive devices and 3D-printed firearms.
The report released this month found that ChatGPT, the most widely used generative AI system, refused only 48 per cent of the more than 2,000 requests entered by testers.
Tech Against Terrorism executive director Adam Hadley said AI models posed a greater risk than a simple web search because they were designed to persuasively engage in back-and-forth conversation.
In Queensland, a 13-year-old boy is alleged to have used AI on multiple occasions before his arrest to simulate a mass shooting.
Police allege the Maryborough boy asked the program to “make me a mass shooting story … kinda like the Bondi beach shooting … a Jewish and blacks festival”.
Adam Hadley says the use of AI by extremists is “ubiquitous”. (Supplied)
A stabbing attack in a Melbourne shopping centre is also alleged to have been the result of AI radicalisation.
Mr Hadley compared the emerging AI platforms to an instructor or mentor capable of giving advice on dangerous topics.
“A model that will iterate with a user, answer follow-up questions and refine an approach is behaving less like a manual and more like a coach,” he said.
“It’s one thing having a bomb-making manual. It’s quite another to have a bomb-making coach at your disposal.“
A spokesperson for the company behind ChatGPT, OpenAI, said it worked “continuously” to make the platform safer by stopping it from providing harmful information.
Tech Against Terrorism found ChatGPT’s safeguards could easily be side-stepped.
The safety features of common AI models are easy to breach, the report found. (ABC News: Greg Bigelow)
The report said urgent action was needed to improve the security of open-source AI models available for download from many developers, including OpenAI, which allowed users to remove the safeguards altogether.
The availability of these models posed “a major national security concern”, the report found.
“We do not accept that developers have an unqualified right to distribute technology that can be used to kill people,”
Mr Hadley said.
“Where a for-profit company releases a capability recklessly, it is right for governments to protect their citizens.”
Extremist use of AI increasing
In his speech this week announcing the national artificial intelligence framework, Prime Minister Anthony Albanese said “extremists and state actors” were known to use AI for propaganda and recruitment.
In 2024 ASIO director-general Mike Burgess told the National Press Club that AI would afford extremists a “step change” in capabilities and warned that offshore extremists were using the technology to build weapons and plan attacks.
Mike Burgess has said AI has become a tool for terrorists. (ABC News: Matt Roberts)
The Tech Against Terror report documented more than 30 cases in which AI systems were used to support attacks by terrorists and extremists. The cases were linked to more than 70 deaths.
Mr Hadley said counterterrorism authorities should place a greater focus on the use of AI by young people for planning and coaching, rather than its use to create propaganda.
“Generative AI … lowers the barrier to entry for precisely the people who are otherwise least capable or most easily deterred: the young, the impulsive and the vulnerable,” he said.
A University of Cambridge study published this month found there was “emerging evidence” terrorist groups and individual actors were using AI to plan violent attacks.
Sam Altman is the chief executive of OpenAI, which is being sued in relation to a deadly university shooting. (Reuters: Shelby Tauber)
OpenAI is under criminal investigation in the United States for ChatGPT’s alleged role in a mass shooting at Florida State University.
Call to geoblock offending platforms
An eSafety Commissioner spokesperson said the agency had “very real concerns” about the growing threat of violent extremists using AI.
Australia’s Online Safety Act requires AI platforms to block access to pro-terror material, including content generated by the AI. Breaches can carry fines of up to $54 million.
University of New South Wales AI Institute chief scientist Toby Walsh said the existing penalties were insufficient given that Anthropic, the creator of the Claude AI, paid $US1.5 billion to settle a copyright lawsuit.
Toby Walsh says stronger penalties are needed to protect Australian AI users. (Supplied: Toby Walsh)
He is calling for fines to be calculated as a percentage of the company’s turnover, or a geoblock that would disable access to unsafe platforms.
“I can’t understand why we’re not more outraged,” Dr Walsh said.
“We can fine them and if that doesn’t seem to work, maybe then we say, ‘Well, sorry, you can’t operate in Australia until you actually follow our laws.
“‘Why should we let you operate here? You’re causing harm to Australian people.'”
