What the letter demands
The open letter contains three main proposals. First, organizations that develop, deploy, and test cyber-capable AI models must immediately adopt high-assurance containment systems. The authors recommend that labs initially run cyber-capable models in a standard mode disconnected from the internet (air gapping).
Second, they propose creating an international body — in consultation with governments and a broad community of experts in artificial intelligence and cybersecurity. This body would standardize best practices for alignment and containment in the public interest, and keep a record of containment breaches.
Third, security incidents related to containment breaches must be reported to this body and ultimately disclosed to the public. According to the letter's authors, such transparency is necessary both to learn from mistakes and to ensure accountability.
The letter also lists short-term measures. Strong containment mechanisms, including operating disconnected from the internet, must be used to reliably hold cyber-capable AI models. For internet-connected models, it notes that the scope of actions permitted by the containment system must be balanced against the current limited ability to predict whether a model's actions will lead to a security breach.
"We, the undersigned scientists, believe that more research into AI safety and greater attention to containment technologies and procedures are both needed." — from the text of the open letter (ai-security-letter.org)
Reuters estimates that on the day of publication, the letter was signed by about 400 scientists from 34 countries. The list on the letter's official website now shows 453 signatories. The Globe and Mail put the number of signatories at more than 260 researchers from 27 countries. According to DigiconAsia, the letter was prepared by University of Toronto professors David Lie, Nicolas Papernot, and Thomas Ristenpart. Among the signatories are AI pioneers Geoffrey Hinton and Yoshua Bengio.
Chain of events: what the letter responds to
The letter was published after a series of security breaches disclosed in recent months. According to DigiconAsia, OpenAI agents were reported to have gained access to systems belonging to the Australian government and to Hugging Face's systems. Anthropic and Google disclosed that breaches were detected in their models during testing.
The letter's text offers two possible explanations for these incidents: the models may not have worked as expected, or the labs may have grossly violated basic cybersecurity rules. In the authors' view, fixes are needed in both cases — alignment research to make models predictable, and stronger containment technologies.
The letter also warns that unless AI development and deployment practices change immediately, such incidents will continue, and future events will cause far greater damage to vulnerable critical infrastructure.
The White House deal and the limits of voluntary commitments
About a month before the letter's publication, Anthropic, OpenAI, Google, Meta, xAI, and Nvidia signed an agreement at the White House on voluntary safety measures. According to Reuters, the agreement includes no mandatory enforcement mechanism.
The scientists' letter focuses precisely on this point: companies' voluntary commitments are not enough. That is why they demand that organizations working with cyber-capable models immediately adopt high-assurance containment standards — until governments develop an official oversight regime.
This event is directly connected to the White House agreement previously covered on our site: Trump and six AI companies signed a voluntary safety deal at the White House. We also wrote on October 10 about the White House requiring AI companies to report incidents mandatorily: The White House required artificial intelligence companies to mandatorily report incidents.
Disagreements: not everyone agrees
The letter was not signed by Meta's chief AI scientist Yann LeCun, who shared the 2018 Turing Award with Hinton and Bengio. In his view, recent security incidents mostly reflect engineering mistakes rather than deeper AI risks.
The EU's tech chief Henna Virkkunen told Reuters that the EU's AI Act is sufficient to address the risk of out-of-control AI systems.
At the same time, the Trump administration on Sunday lifted the ban on US labs giving the UK AI Safety Institute early access to models for safety testing. According to DigiconAsia, Canadian Prime Minister Mark Carney backed the creation of a global AI technology stability board modeled on the Financial Stability Board. One of the letter's authors, Ristenpart, said clear procedural and technical mechanisms are needed to test AI safely.
