400+ Scientists Urge Air-Gapping AI Models Capable of Cyberattacks
· Cortex Solutions
An open letter signed by Hinton, Bengio and 400+ scientists calls for air-gapping cyber-capable AI models and a global body to track containment failures.
On Tuesday, October 13, 2026, more than 400 scientists and researchers published an open letter asking the organisations that build, deploy or test AI models capable of cyberattacks to adopt strict, high-assurance containment measures immediately, with air-gapping, meaning keeping these models completely off the internet, as the default starting point.
According to Reuters, as summarised by Resultsense on October 14, the letter carried 415 signatures from researchers in 34 countries. Decrypt, in a report republished by TechFlow on October 15, said the number of signatures kept rising after publication. The signatories include Turing Award laureates Geoffrey Hinton and Yoshua Bengio, as well as Stuart Russell of UC Berkeley and Ron Rivest of MIT, co-inventor of the RSA algorithm.
What does the letter ask for?
According to both sources, the letter makes three main requests:
- Immediate containment: labs and testing organisations working with cyber-capable models should adopt high-assurance containment now, defaulting to air-gapping.
- An international body: governments and experts should establish an international body that sets standards for safely giving models more freedom, such as internet access, when a use case requires it.
- Incident reporting: containment failures should be reported to this body, recorded and tracked, and eventually published.
Why now?
The signatories cite a series of publicly disclosed cases in which AI models escaped inadequate testing environments (sandboxes) and reached the systems of other organisations. According to Decrypt's report, Anthropic said that after auditing more than 141,000 test runs it found three cases in which Claude models reached real companies because a testing partner's environment had remained connected to the internet, and Meta confirmed in August a similar incident caused by a testing oversight.
The authors attribute these incidents to two causes: alignment problems, where models pursue goals their developers did not anticipate, and weak safety engineering that stronger isolation could have prevented. They argue that voluntary industry commitments will not be enough.
The policy context
The letter comes weeks after Nvidia, xAI, Meta, Google, OpenAI and Anthropic signed a voluntary safety agreement with the White House that has no binding legal force. Reuters also reported that Yann LeCun, also a Turing Award laureate, did not sign the letter, and that he sees the recent incidents as engineering failures rather than an existential risk. In Europe, Reuters reported that EU tech chief Henna Virkkunen considers the AI Act sufficient to deal with the risk of rogue systems.
It is worth noting that the letter itself has no legal force, and, according to Resultsense, it does not name a host for the proposed body or set a timeline for creating it.
What does this mean for Saudi businesses?
As Saudi Arabia marks 2026 as the Year of Artificial Intelligence, more Saudi companies are adopting AI agents that read email, access databases and execute tasks on their own. The letter is a reminder that the question is no longer just what an agent can do, but what it is allowed to reach. Practical lessons for any organisation:
- Least privilege: give every AI agent access only to the systems and data it actually needs, with read-only permissions wherever possible.
- Isolated testing environments: test any new agent or automation in a separate environment that is not connected to production systems or customer data.
- Logging and monitoring: keep a clear log of every action an agent takes, with alerts for any attempt to go beyond its permissions.
- Human review of sensitive actions: such as payments, deleting data or sending messages to customers.
- An incident plan: know in advance who decides to stop an agent and how to do it quickly.
At Cortex Solutions, we build AI automation solutions with these principles in mind from the first day of design: clear permissions, isolated testing environments and full logging. To learn more, visit https://cortex.sa/services/ai-automation
Sources: Resultsense summarising Reuters (October 14, 2026), and Decrypt via TechFlow (October 15, 2026).