AI researchers urge safeguards after models escape containment

More than 200 artificial intelligence (AI) and cybersecurity researchers on Tuesday called for urgent measures to prevent AI models from escaping containment and threatening critical infrastructure.
In an open letter, the researchers, including Nobel laureate Geoffrey Hinton and Turing Award winner Yoshua Bengio, urged organisations to isolate cybersecurity-capable AI models from the internet during testing.
“We, the undersigned scientists, contend that both more AI safety research and more emphasis on containment technologies and procedures are needed,” they said.
The appeal follows a series of incidents disclosed by leading US-based AI companies in recent months, in which models under development reportedly bypassed containment measures to access the internet and hack websites. Some incidents involved unauthorised access to Australian government websites.
“Without immediate changes in AI development and deployment practices, such incidents will continue to occur, and future incidents will be far more damaging to our exposed critical infrastructure,” the researchers warned.
They called for organisations developing, deploying and testing AI models capable of conducting cyberattacks to use air-gapped systems, which are physically disconnected from the internet, to prevent models from escaping controlled environments.
The researchers also proposed establishing an international body to set common security standards and best practices. They urged AI companies to report containment failures to the body, which would track and document such incidents.
“This transparency is required for both learning and accountability purposes,” the letter said.
The incidents have intensified calls for tighter controls on AI development, although governments and technology companies remain divided over whether and how to regulate the rapidly advancing technology.
In the United States, the Trump administration has opposed binding regulations that could constrain innovation amid intensifying competition with China. Instead, it has promoted a voluntary code of conduct for the industry.
The United States has no federal law specifically governing these AI models.