There is a moment in most technology stories when the thing stops being theoretical. When the model scores keep climbing and the benchmarks stop feeling abstract and someone in a boardroom somewhere starts asking, quietly: what happens if this gets used against us?

That moment, for global finance, arrived with the emergence of Anthropic's Mythos model.

Mythos scored 93.9% on SWE-bench Verified and 83.1% on CyberGym. It completed a 32-step autonomous cyber-attack simulation that the UK's AI Security Institute described as a genuine step up from anything that came before. In testing, it solved complex security challenges in 3 out of 10 attempts without any human involvement at all. That last number sounds modest until you consider what those challenges were: multi-step intrusions through enterprise systems, the kind of work that takes a skilled human team days to plan and execute.

Anthropic restricted Mythos to roughly 40 organizations through something called Project Glasswing. The U.S. government simultaneously designated Anthropic a supply chain risk and began integrating the model into NSA operations. The Federal Reserve briefed bank CEOs. Emergency sessions were held in Hong Kong, Singapore, and South Korea. Hong Kong's HKMA accelerated a new Cyber Resilience Testing Framework. The markets reacted: U.S. software stocks dropped sharply after the debut.

You don't convene all of that over a benchmark score. Something real has shifted.

The asymmetry problem

Cybersecurity has always carried an uncomfortable structural problem. A defender has to secure everything. An attacker only needs to find one way in. That imbalance is not new. What is new is the scale at which AI amplifies it.

Mythos can scan entire codebases, identify vulnerabilities across operating systems and enterprise software environments, and suggest exploitation paths. Reports indicate it has uncovered thousands of previously unknown software weaknesses, the kind of flaws that would take elite human researchers weeks to find. Vulnerability discovery, which once required deep expertise and time, is now something a sufficiently capable model can automate.

The economics of this are uncomfortable. Offensive capabilities are getting cheaper and faster. Defense remains expensive, labor-intensive, and slow. The gap between what a sophisticated attacker can attempt and what a security team can realistically monitor is widening. That is not a cybersecurity problem specifically. It is a structural feature of how AI is being deployed, and finance sits directly in the path of it.

Why finance is the target?

Modern banking is not a monolithic system. It is a patchwork: legacy infrastructure built in the 1980s and 1990s, layered with digital overlays, connected to third-party processors, payment networks, cloud providers, and an expanding web of fintech integrations. Each connection is a potential entry point.

Bank executives who were briefed on Mythos were alarmed not just because the model can find vulnerabilities, but because it can suggest how to exploit them. Coordinated attacks on payment and settlement systems, on the data infrastructure that underpins correspondent banking, on the networks that clear transactions across borders. The potential consequences include market disruptions, liquidity freezes, and the kind of trust erosion that can take years to rebuild.

There is also a contagion dimension that makes this harder to contain than traditional financial risk. A successful attack on one institution or one piece of infrastructure in one country does not stay local. Supply chains, cross-border payment flows, and interconnected balance sheets mean that a disruption in one node spreads. The COVID-19 pandemic gave us a clear demonstration of how localized failures cascade into global crises. A coordinated AI-enabled cyberattack could move faster.

The arms race nobody asked for

The uncomfortable reality is that Mythos is not the end of this story. It is the beginning of a new phase of it.

The banking sector is now looking at a continuous cycle: adopt AI for defense, face attacks powered by more advanced AI, invest in new defenses, repeat. Industry analysts expect this to drive significant capital into cybersecurity firms. It is also expected to reshape what cybersecurity professionals actually do: less manual threat hunting, more supervising autonomous agents that monitor systems in real time.

There is a separate concern about what happens when capabilities like Mythos spread beyond the 40 organizations currently approved to use it. The dual-use nature of the technology is not theoretical. The same model that identifies vulnerabilities for defensive purposes can identify them for offensive ones. Anthropic's decision to restrict access buys time. It does not resolve the underlying tension.

The U.S. government's position on this is, to put it charitably, incoherent. Designating Anthropic a supply chain risk while using the model in NSA operations is not a policy. It is a contradiction that illustrates how far regulatory thinking lags behind technical reality. The same tension plays out globally: regulators are racing to build frameworks for a technology that is already several generations ahead of their planning cycles.

What this means for consumers?

Most of the public discussion focuses on institutions and governments. But the impact on ordinary people is just as significant, and probably less understood.

Digital banking is now the default for most financial activity. Mobile payments, online transfers, digital wallets: these are no longer alternatives to traditional banking, they are banking. That infrastructure depends on the same software layers that Mythos-class models can probe.

AI-enabled phishing is already a different category of threat from the clumsy scam emails of a decade ago. Targeted fraud using behavioral data scraped from breached systems, sophisticated impersonation of banking communications, social engineering at a scale and personalization that would have required large teams to execute before: these are not future risks. They are present ones, accelerating.

Consumers lack the resources and technical knowledge to defend against sophisticated attacks. They rely on institutions to protect them. And institutions are being asked to defend against attack capabilities that are evolving faster than their security architecture.

Trust is the foundation of banking. It is also, historically, among the hardest things to rebuild once it fractures.

Nepal: not isolated, just unprepared

Nepal's financial sector is digitizing quickly. Mobile banking adoption has expanded significantly over the past five years. Digital wallets and payment platforms are increasingly central to how people transact. Core banking systems are connected to third-party vendors, international payment networks, and a growing ecosystem of fintech services.

This is, broadly, a positive development. It also creates exposure that did not exist before.

Nepal does not need to develop Mythos-class AI locally for these risks to materialize. Cyber tools developed elsewhere are deployable anywhere. Vulnerabilities in imported software are vulnerabilities in Nepali systems. Cross-border financial linkages mean that a shock transmitted through global payment infrastructure affects Nepali banks regardless of what happens domestically.

The specific vulnerabilities are not hard to identify. Mobile and internet banking grew quickly; security infrastructure did not always keep pace. Older core banking systems integrated with newer platforms carry hidden seams where vulnerabilities can accumulate. There is a genuine shortage of cybersecurity professionals with the expertise to run enterprise-grade threat monitoring. Regulatory frameworks were built for a different threat environment and have not been updated to address AI-driven risk explicitly. Consumer digital literacy remains limited, which makes social engineering and phishing attacks more effective.

None of this is unusual for a rapidly digitizing economy. Most countries in a similar stage of financial development carry similar exposure. What matters is whether the response to that exposure is proactive or reactive.

What actually needs to happen?

The risk here is not exotic. It is familiar: under-investment in security infrastructure, regulatory frameworks that lag behind technology, limited capacity, and the assumption that because nothing bad has happened yet, the threat is manageable.

On resilience: regular vulnerability assessments of banking systems are not optional in this environment. They are routine maintenance. Simulation exercises, stress testing for cyber incidents, and clear incident response protocols need to be standard operating practice, not emergency responses.

On regulation: Nepal Rastra Bank and the relevant supervisory bodies need to update their cybersecurity guidelines to address AI-driven threats specifically. That means integrating AI risk into supervisory frameworks and pushing for industry-wide coordination on incident reporting and response.

On capacity: there is no way around the fact that skilled cybersecurity professionals are scarce globally and particularly so in Nepal. International partnerships with central banks and financial supervisors in more advanced jurisdictions, combined with structured training programs, are the realistic path to building that capacity.

On consumers: awareness campaigns on basic cyber hygiene matter. So do fraud detection systems robust enough to catch AI-generated attacks. Clear, functional grievance mechanisms for fraud victims are not a luxury. They are what determines whether people continue to trust digital banking after an incident.

A broader shift

What makes this particular moment different from previous cycles of cybersecurity concern is the speed of capability improvement. Mythos will be succeeded by Mythos 2, and then Mythos 3. Each iteration will be more capable. The window in which defensive preparation can get ahead of offensive capability is not infinite.

For global policymakers, the challenge is navigating the innovation-security tradeoff without either stifling the technology or leaving critical infrastructure undefended. For financial institutions, it is continuous investment in systems that are never finished because the threat is never static.

For Nepal, the lesson is more direct: the absence of advanced AI in the domestic economy does not insulate the financial system from a world in which advanced AI is shaping the threat landscape. The digitization that has driven financial inclusion over the past decade also created the attack surface that needs to be defended.

The threat is not coming. It is already here. The question is whether the systems that people depend on every day will be ready when they are tested.

This article was written for informational purposes. The author draws on publicly available reports from global regulators, central banks, and financial institutions. Readers with specific cybersecurity concerns should consult qualified professionals.