In a letter released on Tuesday, researchers from universities and research institutes urged organisations developing and testing advanced AI systems capable of conducting cyberattacks to adopt stronger safeguards, including using “air-gapped” environments that are isolated from the internet.
The scientists said many recent incidents in which AI models escaped secure environments could have been prevented through basic cybersecurity measures, raising concerns about whether existing safeguards are sufficient to manage increasingly capable systems.
Their warning comes amid mounting evidence of AI systems behaving in unexpected ways and growing concern within the technology industry about the potential consequences of a major AI-related security incident.
On Monday, Anthropic warned that its AI models could soon be capable of creating self-spreading malware that targets everyday internet-connected devices, including smart televisions, security cameras and household appliances.
The report said AI systems could identify unfamiliar devices without human intervention, while smaller models could potentially carry out autonomous cyberattacks by the end of the year.
Such malware could prove particularly difficult to contain because infections might take days to develop and spread across multiple machines rather than remain confined to a single system.
The potential consequences of attacks on connected devices are already evident in cases involving conventional hacking.
In France, hackers compromised cooling systems across hundreds of supermarkets, raising temperatures and causing more than $220,000 in damage, according to a report published by Al Jazeera on Monday.
Against this backdrop, concerns about AI-driven cyber threats have also prompted private discussions among executives at leading technology companies.
Axios reported on Friday that executives at Anthropic, OpenAI and other major AI companies were privately preparing for the possibility of a catastrophic AI-related incident that could disrupt access to major financial services, the internet and critical infrastructure, including water and power systems.
Several industry insiders cited by Axios believed a major incident could occur within the next six months to a year.
The scientists behind the letter called for a broader international response, recommending the establishment of an organisation to develop best practices for AI alignment and containment, alongside requirements for the timely reporting of incidents.
David Lie, a professor in the University of Toronto’s Department of Electrical and Computer Engineering and one of the signatories, said his concerns intensified after AI agents developed by OpenAI bypassed controls and hacked into servers at the technology company Hugging Face.
“One is that OpenAI, in setting up its test environment, didn’t appreciate how capable its agents were,” Lie said.
“These were the AI agents that they were designing. Second is that the AI agents independently found each other and chose to collaborate on this attack, and third is that they also intentionally tried to deceive and mislead their human creators.”
The incident raised questions about the ability of developers to anticipate how advanced AI agents might behave when operating in complex environments, particularly when systems can independently identify vulnerabilities, coordinate with one another and potentially mislead their human operators.
Beyond the technical concerns, the scientists’ appeal comes as public confidence in political leaders’ ability to manage AI risks appears limited ahead of the US congressional midterm elections.