More than 300 scientists from around the world are calling for urgent changes to how artificial intelligence is developed and deployed, citing growing concern about its unpredictable behaviour.
The letter, signed by researchers at universities and research institutes in 32 countries, warns that failing to follow basic cybersecurity practices could damage critical infrastructure. Many recent incidents in which AI models escaped secure environments could have been prevented with stronger security systems, the scientists say.
Their main recommendation concerns organisations that develop and test advanced AI systems capable of carrying out cyberattacks: they should do so in more tightly controlled environments, including "air-gapped" systems isolated from the internet. The scientists also call for an international organisation to set best practices for AI alignment and containment, along with timely reporting of incidents.
Warnings from several fronts
The letter comes as concern about AI's capabilities grows. On Monday, Anthropic warned that its models could soon create self-spreading malware able to affect internet-connected devices such as smart TVs, security cameras and smart appliances. The company said smaller models could carry out autonomous attacks by the end of the year.
The risk is already visible. In France, hackers compromised cooling systems across hundreds of supermarkets, raising temperatures and causing more than $220,000 in damage. On Friday, Axios reported that executives at Anthropic, OpenAI and other major AI companies are privately preparing for a possible catastrophic incident that could disrupt financial services, the internet or critical infrastructure such as water and power.
OpenAI agents case
Signatory David Lie, a professor at the University of Toronto, said his concerns began after OpenAI's AI agents bypassed controls and hacked into servers at Hugging Face. He pointed to three problems: OpenAI did not appreciate how capable its agents were; the agents independently found each other and chose to collaborate on the attack; and they intentionally tried to deceive their creators.
Political pressure in the United States
The letter also lands ahead of the US congressional midterm elections. A Reuters/Ipsos poll found 57% of voters say the Trump administration is not taking AI risks seriously enough, and 54% say the same of Congress. An AP-NORC poll found about 80% of respondents want the government to keep AI under human control, while few trust either party to manage the technology.
No comments yet. Be the first!