Hinton, Bengio urge new global body to enforce air-gapped AI training standards

Hundreds of AI researchers have issued a formal warning that current security procedures for training artificial intelligence models are insufficient. They are calling for stricter containment measures and the creation of a new international oversight body to manage best practices in AI development.

What happened

According to CSO Online, 487 AI specialists, including senior academics, published an open letter to governments and the AI community. The signatories argue that existing security protocols fail to adequately protect against risks posed by cyber-capable AI models. They specifically recommend using strong sandboxing mechanisms, including air-gapping, which involves disconnecting training systems from the internet entirely. This measure is intended to provide high-assurance containment for models that could potentially be used in cyberattacks.

The letter includes signatures from prominent figures such as Geoffrey Hinton and Yoshua Bengio. These experts suggest establishing a new international organization to oversee AI safety standards. This body would be set up in consultation with governments and a wide range of AI and security specialists. Its primary roles would include defining safety benchmarks and serving as a central point for reporting containment failures. This reporting mechanism aims to improve transparency, facilitate learning from incidents, and ensure accountability across the industry.

Why it matters

The researchers’ demands extend beyond recent general calls for greater AI regulation. Microsoft president Brad Smith recently stated that AI is less regulated than washing machines or milk, though he offered no specific proposals to fix this gap. Similarly, the UK parliament is examining ways to improve AI security, but progress may be slowed by disagreements among AI businesses. The open letter highlights a growing consensus among technical experts that voluntary industry standards are insufficient without stronger, coordinated containment strategies.

Hinton and Bengio have previously engaged in scientist-led initiatives to warn about AI risks. In October of last year, they signed a separate open letter warning of the dangers associated with developing super-intelligent AI. That previous effort did not appear to significantly slow the rapid development of AI technologies. It remains uncertain whether corporate leaders and boardrooms will heed this latest warning regarding the necessity of air-gapped training environments for cyber-capable models.

What to watch

• Monitor for the establishment of the proposed international oversight body for AI safety.
• Track whether major AI companies adopt air-gapping for their training infrastructure.
• Watch for government responses to the open letter and potential regulatory changes.
• Observe if containment failure reporting mechanisms become standardized across the industry.

Who needs to act first

Teams managing AI training infrastructure should immediately audit their current network configurations. Verify whether systems handling cyber-capable model development are fully isolated from external networks. Implement strict sandboxing protocols to contain potential breaches during the training phase. Prioritize environments with the highest access levels to sensitive data or code generation capabilities. Ensure that logging and monitoring tools are active within these isolated environments to detect anomalies. Coordinate with security leadership to define clear containment standards before expanding training workloads.

What to do and how to stay safe: AI security

• Review current AI training environments to ensure they are disconnected from the internet.
• Implement strong sandboxing mechanisms to contain cyber-capable models during development.
• Establish internal protocols for reporting and analyzing any containment failures.
• Monitor for updates from the proposed international oversight body regarding safety standards.

Step-by-step guide: Phishing Explained: How Attackers Steal Trust and Data

General security guidance from the Patch Gazette newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Who signed the open letter demanding air-gapped AI training?

The letter was signed by 487 AI researchers, including senior academics Geoffrey Hinton and Yoshua Bengio.

What specific containment measure do the researchers recommend?

They recommend using strong sandboxing mechanisms, specifically air-gapping, to disconnect training systems from the internet.

What is the purpose of the proposed new international body?

The body would set safety standards and serve as an organization for reporting containment failures to improve transparency.