On 13 October 2026, Reuters reported an international scientific call for containment standards for cyber-capable AI systems and a registry of significant incidents. Geoffrey Hinton and Yoshua Bengio are among the signatories. The proposal includes initial separation from external networks and independent oversight. It is a request from researchers, not an already established international institution or a new rule in force.
The Canadian Press independently corroborates today’s letter and its call for stronger containment. We do not publish an exact signature count: reports use differing numbers, and SIExpo has not independently counted the original list. Both language versions distinguish the verified event from forecasts about future harm.
Two different questions
Restricting access and reporting incidents address different problems. The former concerns what a system under evaluation can reach. The latter helps establish how and why a defense failed. An organisation can have a detailed account after an event while still having set insufficient boundaries before it. A useful mechanism needs to address both, rather than treating transparency as a substitute for containment.
As an editorial example, an agent performing web tasks could be evaluated against copies of services containing synthetic profiles and data. That permits behavioral assessment without automatically granting access to real outside systems. The test must nevertheless describe its simplifications: a result in the copy establishes neither safety nor danger in every public environment.
From proposal to a working process
A useful registry needs a prior definition of what qualifies for entry. A failed request, an access-policy violation and an actual change to somebody else’s data are different events. Counting them identically could make the statistics difficult to interpret. These are our questions for a future mechanism, not final categories published in the letter.
There must also be a verification process. Who submitted the report, what evidence exists, who confirmed the cause and what remains unknown? Public descriptions should inform without turning details of somebody else’s weakness into a tool for abuse. Detailed technical material may require different access arrangements from the general public account.
NIST describes its AI Risk Management Framework as a voluntary approach to incorporating trustworthiness into AI development, use and evaluation. That context shows that risk management can have an established process without an international regulator. The framework does not demonstrate that the authority proposed today exists, nor does it verify a specific incident.
An organisation considering such reporting would also need to distinguish its own observations from a supplier’s explanation. A supplier might identify a configuration error while an operator observes an unexpected action. Preserving both accounts and the unresolved gap is more useful than forcing an early single narrative. This is an illustrative governance problem, not a newly reported case.
Evidence and responsibility
Our interpretation is that a meaningful change would make safety claims inspectable. Clear test conditions, the evaluated system version and actual action records would be needed. A promise or a polished report alone does not establish the effectiveness of containment. The process should make uncertainty visible, rather than obscuring it behind a confidence label.
We do not present forecasts of more damaging incidents as events that have already happened. We also make no legal conclusion about a particular company’s obligations. Today’s international call is the news; whether and how it becomes an institutional process remains unresolved.
Editorial verification: 13 October 2026, 2026-10-13T16:16:52.923011+00:00 (UTC). The cover is an AI illustration, not a photograph of an actual event.
