An April email to bank technology chiefs, released under access-to-information rules, shows OSFI naming Anthropic’s frontier model as a reason the window for fixing flaws is closing.
Regulators do not usually name products. They write about “emerging technologies” and “advanced capabilities” and leave the reader to work out which vendor they mean. So it is worth noticing that when Canada’s banking watchdog wrote to the country’s biggest lenders in April, it named one: Anthropic’s Claude Mythos.
The email, sent on 29 April by the Office of the Superintendent of Financial Institutions to chief technology officers, chief information security officers, and chief risk officers across Canada’s banks and insurers, was obtained by Reuters through an access-to-information request.
“Advanced artificial intelligence models, such as Anthropic Claude Mythos, significantly compress the timeframe for effective risk mitigation,” OSFI wrote. The bulletin, it said, set out “sound practices that institutions can adopt to enhance the speed and effectiveness of risk identification, mitigation and response”.
Parts of the email were redacted under the Access to Information Act, so what OSFI actually recommended remains partly hidden. What is visible is the diagnosis. The traditional patching cycle assumes defenders get days or weeks between a flaw being discovered and being exploited.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!
Mythos, which cybersecurity researchers describe as extremely capable at finding and exploiting software vulnerabilities, collapses that assumption, and banks run some of the oldest software of any industry on earth.
The timing is tidy. In early April, Canadian bank executives met regulators to discuss Mythos, shortly after US Treasury Secretary Scott Bessent and then-Federal Reserve Chair Jerome Powell convened bank chief executives over the same model.
Three weeks later, OSFI sent its email. The pattern has since repeated at the European Central Bank, at the Bank of England, and in Australia, where ASIC confirmed it was watching the model too.
After Reuters put questions to OSFI last week, the regulator published a public bulletin on generative and agentic AI. The sequencing is not subtle.
OSFI’s formal position is that it does not regulate models. “Our focus is not the technology itself, but how federally regulated financial institutions govern and manage the risks associated with its use,” it said.
That is the standard technology-neutral line, delivered in a document that names a specific technology from a specific company twice.
The regulator’s remit runs from the big six banks to pension funds and insurers, and it is charged with spotting risks arriving from geopolitics, foreign interference, and new technology. It is unusual for all three categories to describe the same thing at once.
Access to Mythos is restricted. Euro zone banks appear to be excluded from it, and the model is distributed through Anthropic’s controlled-access programme, Project Glasswing. The Canadian government has said it has access.
Whether any Canadian bank does is unclear, and none of the ones asked would say. Several referred questions to the Canadian Bankers Association, which said its members had invested heavily in protecting the financial system and comply with OSFI’s requirements on cyber risk management and incident reporting, a sentence that answers a different question.
Canada’s big banks are not passive here. Royal Bank of Canada, TD Bank, and BMO have all set out plans to earn money from AI, moving from pilots into chatbots and internal tooling and reducing what they buy from third-party vendors. Scotiabank, CIBC, and National Bank have disclosed initiatives of their own.
Bruce Ross, RBC’s group head of AI, said in June that Mythos marked a shift in the attack landscape, because exploits can now appear as soon as vulnerabilities are found. “The way we’re dealing with it is building our own AI defences,” he said. “We’ll continue to do that.” The defence, in other words, is more of the thing causing the problem.
The wider anxiety in Ottawa is not really about patches. Prime Minister Mark Carney has compared over-reliance on a handful of frontier models to the concentration risk that preceded the 2008 crisis, which is a large claim for a prime minister to make about a chatbot company. OSFI’s email suggests the regulators have stopped treating it as a metaphor.
Get the TNW newsletter
Get the most important tech news in your inbox each week.