Free tools Windows power users keep installed
One-click scans. No signup required.
Bank of England Governor Andrew Bailey is calling for internationally coordinated testing of frontier artificial-intelligence models before they enter wide circulation. His July 14, 2026 speech and September 30, 2026 article present testing as an immediate way to uncover cyber and resilience risks to financial and other critical infrastructure, and as the evidence base for consistent standards later. They do not announce a binding global rule, treaty or implementation deadline.
What Bailey is proposing
Bailey’s proposal is to test the most capable AI models across borders before broad release, rather than relying only on each country’s domestic checks or on monitoring after deployment. He argues that no country can isolate itself from digital systems, models and attacks that cross national boundaries.
As an Amazon Associate I earn from qualifying purchases.
Begin with public authorities
Testing should start at the level of public authorities and extend to systems that underpin national life. That includes payment networks, financial-market infrastructure, banks and other critical institutions. Public-sector testing would help authorities understand how a model behaves against real operational environments before firms widely integrate it.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Test critical infrastructure, not just model outputs
The concern is not limited to whether a model gives an incorrect answer. Bailey says cutting-edge models could identify and potentially exploit weaknesses in information-technology systems. Evaluations therefore need to examine how models interact with defensive controls, permissions, networks and recovery procedures, including the possibility of deliberate misuse.
Pair model testing with recovery capability
Bailey’s resilience response includes stronger recovery arrangements. Firms should be able to restore essential systems after a severe compromise, including “Bare Metal recovery”: rebuilding systems from scratch rather than assuming that existing software, configurations or data remain trustworthy.
Why the Bank of England sees this as a financial-stability issue
Financial institutions depend on tightly connected technology. A vulnerability found in one provider or market infrastructure can spread through payment, settlement, communications or data links. Frontier models may accelerate vulnerability discovery, automate parts of an attack or behave in ways developers did not anticipate. An incident affecting a major institution could therefore become an operational and confidence shock, not merely an isolated software defect.
Rank #2
International coordination matters because the same model may be trained in one jurisdiction, hosted in another and used by institutions around the world. A purely national test could miss cross-border dependencies and leave firms subject to inconsistent expectations. Bailey’s argument is for a common testing direction, while recognising that implementation would still involve national authorities and regulated firms.
Recommended Free Tools
How the proposed approach would work
- Evaluate before broad circulation. Test a frontier model before it is made widely available or embedded in critical services.
- Use public-authority scenarios. Authorities and infrastructure operators would examine the model in environments representative of essential services.
- Probe for cyber and operational weaknesses. Assess whether the model can expose vulnerabilities, bypass controls, create unsafe actions or produce unexpected behaviour.
- Record incidents and near misses. Failed tests, harmful outputs and events that almost caused damage would become evidence rather than being treated as isolated anecdotes.
- Feed evidence into common standards. Over time, authorities could codify what reliable testing, intervention and recovery should look like across finance and potentially other sectors.
- Maintain recovery plans. Organisations would need tested procedures for containment, restoration and bare-metal rebuilding when prevention fails.
Bailey’s September article makes the rationale explicit: “Testing will not eliminate failures. Models will behave unexpectedly. They will expose weaknesses in our assumptions. That is not evidence that testing has failed, it is evidence of why testing is necessary.”
Rank #3
How this differs from other policy choices
| Policy question | Bailey’s emphasis | What it would address |
|---|---|---|
| When to act | Pre-release testing, followed by monitoring after deployment | Unknown capabilities before they reach critical systems, plus changing risks in use |
| Geographic scope | International coordination rather than isolated national checks | Cross-border models, suppliers, infrastructure and attacks |
| Who tests | Public authorities working with firms and infrastructure operators | Risks that a vendor’s own evaluation may not reveal in public-service settings |
| What gets tested | Model behaviour, cyber exposure, intervention points and operational recovery | Both the AI system and the surrounding organisation’s ability to withstand failure |
| Legal force | Testing as a practical step toward later standards | Evidence for possible consistent requirements; no binding global rule is announced |
Has a global AI testing standard been adopted?
No. The July and September 2026 Bank of England materials describe a policy call and an emerging approach, not a completed international standard. They do not establish a universal test protocol, a treaty, mandatory participation or a deadline for implementation.
Bailey places testing before a more formal regulatory framework because real evaluations can reveal unexpected behaviour and credible points where authorities or firms can intervene. Incidents and near misses could then help turn practical lessons into consistent requirements. He says the United Kingdom is well placed to contribute through institutions including the AI Security Institute and the National Cyber Security Centre, but that does not amount to a global regime.
What firms and infrastructure operators should take from the proposal
- Map AI dependencies: identify where external or internal models influence payments, trading, fraud controls, customer operations or security decisions.
- Demand operational evidence: ask vendors how models were evaluated for misuse, unexpected actions and interactions with real systems.
- Run adversarial exercises: test permissions, network segmentation, monitoring and human approval points under hostile or abnormal model behaviour.
- Practice restoration: maintain offline, trustworthy backups and rehearse bare-metal rebuilding rather than relying on an untested recovery document.
- Report near misses: preserve logs and lessons from events that were contained before causing material harm; those cases are valuable for future standards.
These are implications of Bailey’s resilience argument, not a new Bank of England rule announced in the cited speeches.
Why Bailey links AI safety to economic growth
Bailey has also described AI as a likely next general-purpose technology. In a May 21, 2026 speech, he contrasted average annual UK potential-supply growth of 2.8% before the financial crisis (1990–2006) with 1.3% afterward. UK national income per head grew by 2.0% a year before the crisis and 0.6% afterward. Those figures explain the balance in his position: AI could support productivity and growth, but unmanaged failures in widely connected systems could impose large economic costs. Testing is presented as a way to pursue the opportunity while improving resilience, not as a proposal to halt innovation.
The unresolved questions
A coordinated programme would still need agreement on which models qualify as “frontier,” who has access to confidential test results, how cross-border investigations are governed and what happens when a model fails. It would also need to distinguish laboratory capability evaluations from tests of a deployed system’s people, permissions and infrastructure. Bailey’s speeches identify testing and recovery as starting points, while leaving those institutional and legal details for subsequent policy work.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
