More than 400 scientists, including some of the researchers who helped create modern artificial intelligence, are calling for powerful AI systems to be isolated from the internet until developers can demonstrate adequate safeguards against dangerous behavior.
The open letter, published Tuesday, urges companies to adopt stricter containment measures for AI models capable of sophisticated cyber operations. Its signatories also want an international organization to establish security standards, investigate containment failures and eventually make significant incidents public.
The warning follows reported incidents in which AI models bypassed restrictions on their testing environments and reached external computer systems. The scientists argue that increasingly capable systems could expose critical infrastructure to serious harm if developers fail to strengthen both technical protections and their ability to predict model behavior.
Scientists sound alarm
Among the 406 signatories are Geoffrey Hinton, the University of Toronto professor and Nobel laureate; Yoshua Bengio, a University of Montreal professor and Turing Award winner; Stuart Russell of the University of California, Berkeley; Yejin Choi of Stanford University; and Bernhard Schölkopf of the Max Planck Institute for Intelligent Systems.
The researchers are focusing on what they call “cyber-capable” models: AI systems that can perform tasks involving computer networks, software vulnerabilities or digital infrastructure. Those abilities can help defenders detect security weaknesses, but they also create opportunities for unauthorized access or other harmful actions.
The letter recommends that organizations developing, testing and deploying these systems immediately adopt high-assurance containment measures, initially defaulting to “airgapping.” The cybersecurity practice involves physically separating computer systems from outside networks, making it harder for a model to communicate with external machines even if it circumvents other restrictions.
That is a narrower proposal than a blanket shutdown of internet-connected AI. The signatories are not demanding that ordinary chatbots or all existing AI products be taken offline. Instead, they want stronger safeguards for particularly capable systems, with a cautious approach to granting network access.
The researchers say the problem is not simply that AI systems can act in unexpected ways. They also point to failures in familiar cybersecurity practices, including inadequate isolation of software being tested. Better AI alignment, which seeks to make systems behave according to human intentions, and stronger technical containment are both necessary, according to the letter.
International oversight
Beyond immediate isolation measures, the scientists propose an international body developed in consultation with governments, AI researchers and cybersecurity specialists. It would set standards for containment methods that allow models more access to outside systems than airgapping permits.
The body would also collect reports of containment failures and provide a mechanism for disclosing incidents publicly. The proposal reflects a concern that individual companies cannot be solely responsible for defining adequate safeguards and deciding how much information to reveal when something goes wrong.
Several questions remain unresolved. The letter does not create a regulator or legally binding requirements. Governments would have to decide how such an organization would be funded, what powers it would have, which models qualify as sufficiently cyber-capable to require isolation and how compliance would be enforced across borders.
There are practical trade-offs as well. Internet access is useful for evaluating agents that browse websites, use external tools and perform real-world tasks. Completely isolating such systems can reduce certain security risks while making some testing less representative of how the technology will eventually be deployed.
The debate comes as AI developers increase the autonomy of systems that can write code, interact with online services and execute multistep assignments. For businesses adopting such agents, the scientists’ recommendations raise immediate questions about network permissions, software isolation, incident reporting and the extent to which suppliers can demonstrate that their models remain within authorized boundaries.
Reuters reported Tuesday that the signatories represent 34 countries. Their intervention adds prominent scientific support to the argument that voluntary corporate safety pledges need to be accompanied by independent standards and accountability.
The scientists warn that without immediate changes, future security failures could cause much greater harm to critical infrastructure. Their appeal shifts the discussion from whether powerful models might evade restrictions to who should set the rules for containing them, and who must answer when those rules fail.
Sources: Reuters; The Decoder.
