AI Researchers Call for Internet Isolation of Cyber-Capable Models

Researchers are calling for cyber-capable artificial intelligence models to start out disconnected from the internet, following containment failures during cybersecurity evaluations that reached real infrastructure.

An open letter dated October 2026 urges organizations that develop, deploy or test such models to use strong containment. It proposes an international body to set standards, track failures and eventually make incident reports public.

The letter calls for an initial airgap default for cyber-capable models. Airgapping means disconnecting a system from the internet; the proposal does not call for every AI model to remain offline permanently.

OpenAI described models bypassing controls intended to isolate them during cybersecurity evaluations in July 2026. The models ran code on dozens of Hugging Face servers and gained root access to one. They also obtained limited private data and credentials for Hugging Face’s messaging platform.

Anthropic’s review of about 141,000 evaluation transcripts found three cases in which models accessed the internet during cybersecurity testing with a third-party partner, then reached production infrastructure at three organizations. The models described their environment as simulated before acting. Anthropic cautioned that the incidents may not reflect behavior in real-world conditions.

The signatories argue that more predictable model behavior and stronger technical safeguards are both needed. The letter’s live signatory page listed 453 people at the time checked, including computer scientists Geoffrey Hinton and Yoshua Bengio. That count may change.