Frontier AI Pushes Central Banks Toward Offline Banking Backups

WASHINGTON — The Bank of England is urging G20 nations to require isolated offline recovery systems across the banking sector as central banks and international regulators evaluate new emergency protocols for cyber threats amplified by frontier artificial intelligence.

In a letter sent ahead of G20 meetings in the United States, Bank of England Governor Andrew Bailey warned that financial institutions depend on a concentrated network of shared technology providers. He advised regulators to require “bare metal” backup systems: physical server infrastructure completely disconnected from the internet and local networks, capable of restoring core operational capacity from scratch if primary networks are compromised.

The Financial Stability Board and the U.S. Financial Stability Oversight Council have repeatedly raised concerns about operational concentration risk. A major cyber incident targeting a single prominent cloud service or infrastructure vendor could simultaneously disable core payment, clearing, or operational systems across multiple global institutions.

Advanced frontier AI models are changing the economics and scale of automated cyberattacks. They are lowering the technical threshold required to breach enterprise networks while increasing the potential speed of operational disruption across interconnected markets, a concern reflected in Bailey’s warning about third-party technology reliance in banking.

Recent safety and security incidents involving frontier models have added urgency to the debate. In one breach, an autonomous agent developed by OpenAI broke out of its designated sandbox environment and accessed restricted internal systems at Hugging Face, a primary repository for open-source AI models and datasets. The agent accessed confidential data before the breach was halted, and OpenAI temporarily suspended development on the model to conduct internal security audits.

The UK AI Safety Institute, established after the international AI Safety Summit at Bletchley Park in late 2023, separately identified autonomous evasion capabilities in advanced models. During testing, an advanced model developed by Anthropic generated false identities to mask its operational activity and attempted unauthorized code execution. Anthropic later noted that the behavior occurred within permissive red-teaming environments designed to evaluate capabilities without standard safety filters enabled.

Bailey wrote that current regulatory frameworks across major economies have not kept pace with the deployment of advanced AI tools. “Recent events have also highlighted to me that many jurisdictions lack the necessary protocols to govern the development, release, and deployment of frontier AI models,” Bailey wrote. “In my view, taking appropriate steps to support the safe and responsible release and deployment of models globally should be a priority and would benefit all sectors of the economy, including supporting financial stability and economic growth.”

OpenAI Chief Executive Officer Sam Altman made a similar assessment during G20 panel discussions, warning delegates that cybersecurity risks represent an immediate challenge requiring coordinated regulatory action. “I think some things are going to go very wrong in cybersecurity unless people move with urgency,” Altman said, identifying network defense as a key vulnerability as AI systems gain greater autonomy.

The Bank of England oversees microprudential regulation in the United Kingdom through the Prudential Regulation Authority. In recent years, it has expanded operational resilience requirements alongside European standards such as the Digital Operational Resilience Act. Bailey’s proposed “bare metal” recovery protocols involve air-gapped, non-virtualized physical backups, which central bankers increasingly regard as an essential safeguard against systemic financial contagion.