Anthropic's Claude Mythos and the Regulatory Reckoning: When Frontier AI Becomes a Financial Stability Risk
The BoE, FCA, Treasury and NCSC convene emergency talks with UK banks after Anthropic's new model discovers thousands of previously unknown software vulnerabilities, including some undetected for decades.
Context:
In mid-April 2026, the Financial Times reported that the Bank of England, Financial Conduct Authority, HM Treasury, and the National Cyber Security Centre were holding urgent, coordinated discussions with major UK banks, insurers, and financial infrastructure providers through the Cross Market Operational Resilience Group (CMORG). The trigger: Anthropic's new AI model, known as Claude Mythos Preview, had demonstrated an unprecedented ability to autonomously identify and exploit zero-day vulnerabilities across major operating systems, web browsers, and enterprise software.
Anthropic has stated that Claude Mythos Preview exceeds 'all but the most skilled humans' in detecting software vulnerabilities. The model has already, according to the company, discovered thousands of previously unknown security bugs, including some that had remained undetected in critical infrastructure for decades. Anthropic launched 'Project Glasswing,' a controlled access coalition of over 40 major technology and financial companies (including Amazon, Apple, Google, Microsoft, Cisco, Nvidia, and Broadcom) to study the model's implications for cybersecurity before wider deployment. This is the first time Anthropic has imposed such restrictions on any of its products.
Bank of England Governor Andrew Bailey publicly stated that cybercrime risk must be reassessed in the context of this development. Barclays CEO C.S. Venkatakrishnan warned that banks “have to worry”. Finance ministers attending the IMF Spring Meetings in Washington pressed peers on the Mythos model. US Treasury Secretary Scott Bessent convened American bank executives for an equivalent briefing, an unusual coordination between London and Washington that signals the seriousness with which senior policymakers are treating the systemic risk dimension. UK banks were given controlled access to Mythos from the week of 20 April, allowing them to begin testing their own systems for the vulnerabilities the model can identify.
Rules and Guidelines:
The legal and regulatory framework in which CMORG is operating is the UK's operational resilience regime, specifically the FCA and PRA Policy Statement PS26/2 (published March 2026) sets out final rules and guidance for reporting operational incidents and material third-party arrangements.
CMORG's mandate is to coordinate the operational resilience response across the UK financial system. Its membership includes the Bank of England, NCSC, FCA, HM Treasury, UK Finance, eight systemically important UK banks, four financial infrastructure providers, and two major insurers. Its activation in the Mythos context is significant: it is the mechanism through which a potential systemic cyber risk, one that does not originate from a single firm's failure but from a capability change in the external threat environment, is assessed and communicated across the sector.
The dual-use nature of Claude Mythos is the central regulatory challenge. The same capability that enables Mythos to discover and patch vulnerabilities faster than human security teams can also be used offensively. Anthropic has been explicit that Mythos can both identify and exploit vulnerabilities, making it a potential offensive weapon in the hands of a malicious actor who obtains access. The FCA and PRA's existing cyber resilience supervisory expectations were designed for a threat landscape in which human adversaries were the primary risk. There is also the question of concentration risk: Mythos's attack surface is not any single bank but every operating system, browser, and enterprise tool used across the sector's technology perimeter.
Businesses Affected:
All UK-regulated banks, building societies, insurers, and financial market infrastructure providers whose entire technology perimeter is in scope of the vulnerabilities Mythos can identify.
Third-party technology and cloud providers serving the financial sector whose products and infrastructure are implicated in the vulnerabilities Mythos has identified, and who have received controlled access through Project Glasswing.
Legal and compliance teams managing PS26/2 implementation must assess whether the Mythos development constitutes a material change to their firm's third-party risk assessments and whether any notification obligations arise.
Next Steps:
UK banks should expect formal briefings from CMORG in the coming weeks on the specific categories of vulnerability Mythos has identified. Attend these briefings and ensure your CISO and operational resilience team lead are represented.
Request access to the controlled Mythos testing programme. The FCA and NCSC are facilitating UK bank access for defensive testing. Firms that use this access to identify and patch vulnerabilities before a malicious actor exploits them will be materially better positioned than those that do not.
Conduct an urgent third-party dependency audit. The key question is whether every operating system, browser, and enterprise tool used across your technology perimeter has received recent security updates. Given Mythos has identified vulnerabilities undetected for decades, assume your current estate contains unknown exposure.
Review PS26/2 third-party reporting obligations in light of the Mythos risk. If any of your material third-party providers have been notified of Mythos-identified vulnerabilities that have not yet been patched, assess whether this constitutes a material third-party arrangement change requiring regulatory notification.
Brief your board and risk committee on the Mythos development as a standalone agenda item. This is not a routine cyber update, it is a structural shift in the threat landscape with potential financial stability implications. Board-level awareness and response planning is a regulatory expectation under SM&CR and operational resilience rules.
Source | FT | Operational/Cyber Risk
