An international group of scientists – including leading figures in AI – have signed a letter urging for the regulation of advanced artificial intelligence models by a new international body.

The scientists, which including AI leaders Geoffrey Hinton and Yoshua Bengio, are calling for a new organisation to standardise containment strategies and incident reporting for AI systems with advanced AI capabilities.

The letter, signed by around 400 scientists, comes after several high-profile incidents of breaches by new AI models from OpenAI, Anthropic, Google, and Meta.

The newly propose organisation would aim to set containment standards for the AI industry in an effort to prevent such instances of breaches.

It would also establish a register of security events perpetrated by AI models, which would eventually be made public, the scientists said.

The letter urges that voluntary commitments – such as the ones made by top AI firms to the White House – on AI safety are not enough to secure frontier AI models or ensure their testing is truly safe and secure.

Specifically, the letter proposes that frontier AI models should be completely compartmentalised from external networks, and should be under the review of independent oversight, in order to avoid past mistakes in sandboxing which allowed AI models to infiltrate and breach different external systems.

It urges organisations currently testing cyber-capable AI models to adopt high-assurance containment standards.

The letter comes as the UK’s AI Minister warns that heightened AI cyber capabilities – or merely the testing of new AI models – would “likely” breach the UK’s cyber defences at some point.

Recommended reading

• Microsoft Reveals How Quickly AI Is Weaponising Vulnerabilities
• ICO Secures Better Data Protection from AI Developers
• OpenAI Firings Highlight Wider AI Security Concerns

Kanishka Narayan warned that a UK government cyber breach was “likely, if not inevitable” in the wake of OpenAI’s AI model hacking of the Australian government’s health system.

He gave his own message to AI developers: “Listen when your scientists warn of serious dangers. Investigate unexpected behaviour. And if you cannot adequately control the risks, do not release the system.”