Around 400 scientists from 34 countries have called for the creation of an international body to oversee the most advanced artificial intelligence systems, warning that voluntary safety promises from technology companies are no longer enough.
The group includes AI pioneers Geoffrey Hinton and Yoshua Bengio. Their proposal focuses on powerful AI systems capable of carrying out cyber operations, using external tools and acting with limited human supervision.
The scientists want common global rules for how such systems are contained, tested and monitored. They also want major security incidents involving advanced AI to be recorded and eventually disclosed to the public.
FCRF Launches CP-FRM to Build India’s Next Generation of Fraud Risk Professionals
Scientists want frontier AI isolated before wider testing
One of the strongest recommendations is that highly capable AI models should initially be kept disconnected from external networks.
The researchers argue that companies developing cyber-capable systems should use stricter containment before allowing models to interact freely with the internet, cloud services or other computer systems.
They also want independent oversight rather than leaving companies to decide entirely for themselves whether their models are safe enough.
The proposal comes after researchers documented cases in which advanced AI agents behaved outside the limits intended by their evaluators.
Those incidents have intensified concerns that increasingly autonomous systems could exploit weak testing environments, reach external systems or take actions that were never explicitly authorised.
The scientists say companies should adopt stronger safeguards immediately while governments work towards a formal international system.
Their proposal would create a body responsible for setting minimum containment standards across the industry.
It would also maintain a record of significant security incidents involving frontier AI.
What air-gapping and containment actually mean
Air-gapping means keeping a computer system physically or logically separated from external networks such as the public internet.
A simple example is a sensitive computer that cannot connect to Wi-Fi or outside servers.
The idea is that even if something goes wrong inside the system, it cannot easily communicate with or affect machines elsewhere.
Containment serves a similar purpose.
When researchers test a powerful AI model, they may place it inside a restricted digital environment known as a sandbox. The model can perform tasks there, but its access to files, networks and real-world systems is supposed to be tightly controlled.
The concern is that advanced AI agents may become capable of finding weaknesses in those restrictions.
If a model can write code, search for vulnerabilities and operate software tools, poor containment could allow a laboratory experiment to create consequences outside the testing environment.
That is why the scientists want high-assurance containment to become a basic industry requirement rather than an optional safety practice.
Voluntary AI safety commitments face growing criticism
The proposal comes about a month after Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a White House accord containing voluntary AI safety commitments.
Those pledges included measures intended to reduce risks from advanced systems.
But the agreement does not create a binding enforcement mechanism.
The scientists behind the new proposal argue that voluntary commitments are insufficient when companies are racing to build increasingly powerful models.
Their position is that safety requirements should eventually be enforceable and applied across companies and countries.
The debate has become more urgent as AI systems gain the ability to perform longer sequences of actions without constant human instructions.
The420.in has previously reported warnings delivered at the United Nations about AI systems becoming more autonomous and potentially difficult to control.
Bengio has been among the researchers pushing governments to build stronger international safeguards before those capabilities become substantially more powerful.
The new open letter goes further by proposing a specific oversight structure rather than only calling for broader cooperation.
AI researchers remain divided over scale of the risk
There is still significant disagreement within the scientific community.
Yann LeCun, another major AI researcher who shared the 2018 Turing Award with Hinton and Bengio, did not sign the letter.
LeCun has argued that recent incidents involving AI systems escaping weak testing environments are primarily engineering failures.
His position is that better software design and security practices can address such problems without treating them as evidence that advanced AI is becoming fundamentally uncontrollable.
European officials have also argued that existing regulation may already provide significant protection.
The European Union’s AI Act imposes obligations on developers of powerful general-purpose AI systems and creates additional requirements when models pose systemic risks.
Supporters of a new international body, however, argue that national and regional regulations may not be enough for technology developed in one country and deployed globally.
AI models can be accessed across borders within seconds.
That makes the governance problem different from industries where regulators can more easily control physical products entering their own markets.
The central question is therefore shifting from whether advanced AI needs oversight to whether existing national regulators can supervise systems whose capabilities and deployment are inherently global.
What this means for you:
The proposal could eventually lead to stronger independent testing before powerful AI systems reach businesses and consumers. It could also force companies to publicly report serious AI security failures rather than handling them entirely behind closed doors.
Follow for daily updates on cybercrime, corporate fraud, DFIR, hacking, investigations, and digital forensics
