More than 200 experts urge isolating cyber-capable AI models from the internet — NDTV World

More than 200 researchers in artificial intelligence and cybersecurity have called for immediate changes to approaches to developing and testing AI capable of performing cybersecurity tasks. They propose physically separating such models from the internet to prevent them from leaving the testing environment, NDTV World reports.

Requirement for physical isolation

The letter was signed, among others, by Nobel Prize laureate Geoffrey Hinton and Turing Award laureate Yoshua Bengio. The authors of the appeal said that AI safety research must be strengthened and that more attention should be paid to technologies and procedures for containing models.

The scientists called on organizations that develop, deploy, or test AI models with cyber capabilities to apply air-gapping. This involves physically disconnecting systems from the internet to prevent unauthorized access by models to the network.

More current news is available on the UA.News Telegram channel Telegram.

Reports of containment failures

Leading U.S. AI companies have in recent months disclosed cases in which models at the development stage escaped containment measures, gained internet access, and hacked websites. Such cases included unauthorized access to Australian government websites.

The signatories warned that without immediate changes to AI development and deployment practices, such incidents would continue, while future breaches could cause significantly greater damage to critical infrastructure. They also proposed creating an international body to standardize security practices and requiring companies to report model containment failures to it.

The appeal emphasized that transparency is needed to learn from such incidents and ensure accountability. The publication notes that the United States has no specific federal law directly regulating such models, while Donald Trump's administration supports a voluntary code of conduct instead of mandatory rules.