Share the article
Reg rageJuly 3
Politically sensitive frontier AI is forcing global banks to decide who can use it
by
Share the article
© Andrey Rudakov/Bloomberg
The most dangerous sales pitch in financial technology is no longer proving a tool is cheaper, faster or more efficient. It is in establishing it is vastly better
That is the lesson from the recent scrutiny around Anthropic’s frontier models and their use by global banks. JPMorgan has reportedly stopped Hong Kong staff accessing Anthropic’s models, following similar caution elsewhere, amid rising concern over where US-origin frontier AI can be used and by whom
This is a problematic development. Some AI tools are now being sold on capabilities so powerful that they cannot plausibly remain ordinary vendor products
Take cyber-capable AI. Anthropic’s Project Glasswing says its Mythos Preview model has been used to find hundreds of high- or critical-severity vulnerabilities in open-d effort: find weaknesses, fix them faster, make critical systems safer. That is a reasonable ambition, but it is also precisely why governments get nervous. In cyber security, the side that finds the weakness first has the advantage
Fomo drives poor decision-making
For banks, the attraction is obvious. Their technology estates are vast, old, patched, interconnected and systemically important. A model that can scan code, identify flaws and accelerate remediation is the sort of product no chief information security officer wants to ignore. If a rival can use AI to harden its infrastructure faster, the pressure to adopt becomes intense. AI procurement is no longer driven only by productivity fear of missing out. It is becoming defensive Fomo
But defensive Fomo is still Fomo. And in heavily regulated finance, that matters
The danger is that banks buy advanced AI as if it were software-as-a-service, when some of these tools increasingly resemble critical infrastructure, dual-use technology and geopolitical chokepoints. Can a bank safely depend on a tool whose best commercial argument is also the reason Washington, or any other government, may restrict it?
That is the Mythos paradox. The vendor says: this model can find serious vulnerabilities. The regulator hears: this model can find serious vulnerabilities. Same sentence; opposite reaction
The recent US intervention over Anthropic’s Fable and Mythos models makes this procurement risk concrete. Anthropic said the US government issued an export-control directive suspending access to the models by foreign nationals, including foreign-national employees. While Anthropic disputes the government’s rationale, the lesson for banks is that access to frontier AI can be reshaped overnight by national security decisions far outside typical vendor-management process
Your nationality matters
That makes global governance ugly. The practical result is potentially two-tier internal friction inside institutions that are supposed to operate globally. A model may be available to one desk, blocked for another, usable in one jurisdiction, restricted in another, approved for some employees and unavailable to others. The same bank then has to maintain different AI toolchains, different approval processes, different audit trails and different assumptions about what staff can safely do
If a global bank cannot apply the same cyber defence process across its own estate, supervisors will want to know how consistent, auditable and resilient that process really is. And if many banks respond by clustering around the same small set of approved frontier models, the issue becomes a concentration and operational risk to boot
The vendor’s incentives are not reassuring. Frontier AI companies, like Anthropic, need enterprise customers. They must prove that their models are indispensable systems. The way to do that is to demonstrate increasingly consequential capabilities, yet every step up the capability ladder strengthens the case for regulatory scrutiny
This is the trap. If the tool is only modestly useful, banks may not need it. If it is transformative, governments may not leave it alone
Banks should treat frontier AI tools in sensitive domains as politically exposed infrastructure with board-level recognition that some vendor products now carry export control and national security risk
For years, banks have been told that AI adoption is a competitiveness test. That remains true. But the Mythos episode suggests a second test is arriving. Can banks distinguish between ordinary automation and capabilities that are so powerful they become matters of state?
The answer will determine whether financial institutions use AI to reduce risk or build a new dependency they do not control
Was this article helpful?
Thank you for your feedback!
Your feedback will help us improve our article content
Trending in The Banker
-
Standard Chartered loses bid to have 1MDB trial dismissed
Crime & ConductJULY 1
-
UK court rules against banks in motor finance case
Car loan scandalJUNE 30
-
‘To be clear, Close Brothers Limited has not filed for administration’
ExclusiveJUNE 24
-
TSB staff ordered back into office under Santander
ExclusiveJUNE 30
-
Barclays buys London HQ for £750mn
UKJUNE 30
