AI is starting to find the flaws in banks: global regulators are stepping in
Artificial intelligence has long been presented as a writing machine: it drafts, summarises, translates, produces images and helps with code.
But another stage is arriving. A far less comfortable one.
AI systems are becoming able to search for flaws in the computer systems that companies, banks and critical infrastructure depend on.
And this time the subject is not staying in the labs or at cybersecurity conferences. It is landing on the desk of global financial regulators.
On 18 May 2026, Reuters and The Guardian report that Anthropic is to brief the Financial Stability Board — the international body that coordinates financial rules across the G20 — about its Claude Mythos Preview model.
Why would an AI model interest financial regulators?
Because it is not only there to talk. It can help find software vulnerabilities. And what helps defenders can also, if poorly controlled, help attackers.
AI no longer merely threatens to produce bad text.
It is starting to touch the systems that keep money moving.
What you need to understand in 30 seconds
Claude Mythos is not one more consumer chatbot.
Anthropic presents it as a model that performs particularly well on cybersecurity tasks. Its official purpose is defensive: helping certain authorised actors identify and fix flaws in critical software and systems.
But that is precisely what makes the subject sensitive.
When an AI becomes strong enough to look for flaws, the same progress has two faces:
- on the defensive side, it can speed up patching vulnerabilities;
- on the offensive side, it can cut the time, cost and expertise needed to find weak points;
- on the financial side, it can turn a technical problem into systemic risk.
That is why the subject goes beyond Anthropic. It raises a broader question: what happens when the best AI models start automating part of the work of cybersecurity experts?
The timeline that raises the alert level
This did not come out of nowhere. It built up over a few weeks.
7 April 2026 — Anthropic unveils Claude Mythos Preview.
The company describes a general-purpose model that performs very well on cyber tasks, and announces Project Glasswing, a controlled-access programme aimed at helping certain actors secure critical software and systems. Two months later, a guarded public version of that same family shipped under the name Fable 5 — we looked at what Anthropic opened up and what it deliberately held back.
13 April 2026 — the UK AI Security Institute publishes its evaluation.
In controlled tests, Claude Mythos Preview shows a capability jump, notably on multi-step attacks against vulnerable networks.
7 May 2026 — the IMF sounds the alarm.
The International Monetary Fund warns that AI can amplify cyber threats to the point of becoming a financial stability risk.
13 May 2026 — another warning from the United Kingdom.
The AI Security Institute reports that the autonomous cyber capabilities of frontier models are advancing very fast, on a scale of months rather than years.
18 May 2026 — the subject goes global.
Reuters and The Guardian report that Anthropic is to brief the Financial Stability Board on the risks linked to Mythos.
That timeline matters. It shows this is not a vague fear about “dangerous AI”. It is a rapid, documented escalation that is starting to interest international financial authorities.
From assistant to flaw-hunter
For the general public, a cyberattack often looks like a film scene: a hacker in the dark, green lines of code, a password cracked in ten seconds.
Reality is less spectacular, and more worrying.
A successful attack often comes from a chain of small weaknesses: an unpatched piece of software, a bad configuration, an old vulnerability, a legacy system kept because it still works, a forgotten door in a complex infrastructure.
That is exactly where AI can change things.
It does not necessarily replace the brilliant hacker. It can replace the time.
It can search faster, test more leads, read more code, compare more behaviours and spot combinations humans would take far longer to explore.
So the danger is not only that AI invents new attacks. The danger is that it industrialises the search for existing flaws.
And in finance, existing flaws are never a mere technical detail.
Why banks are an explosive case
Banks, insurers, financial service providers and market infrastructure run on complex systems, often old, often interconnected.
There is legacy software.
Shared providers.
Technical dependencies.
Deferred patches.
Critical systems few people fully master.
Software chains shared between several players.
As long as finding flaws takes a lot of human time, part of the system holds together thanks to that friction.
But if AI reduces the friction, the balance of power changes.
What was hard to find can become more accessible. What needed a team can need fewer people. What took weeks can take much less.
This is not the end of the world.
It is more awkward than that: it is a credible, gradual, concrete problem, and it is already being discussed by financial authorities.
The IMF is not talking about a mere IT problem
On 7 May 2026, the IMF publishes a clear warning: AI can amplify cyber threats, and those threats can become a financial stability problem.
That is not only a technical alert.
A major cyber incident can trigger funding stress, solvency worries, market disruption and a loss of confidence.
In finance, confidence is not a moral extra. It is the operating system.
If an AI makes it faster to find flaws in financial systems, the risk no longer concerns an IT department. It concerns the infrastructure that moves money.
What 18 May really changes
The news of 18 May 2026 does not mean Claude Mythos will attack banks tomorrow morning. That would be false and sensationalist.
What it means is more serious: the subject is entering the radar of global financial regulators.
The Financial Stability Board is not a small technical committee. It brings together officials from central banks, finance ministries and regulators across the major economies.
When a body like that takes an interest in an AI model, it means the risk is no longer treated as a laboratory matter.
It becomes a financial stability issue.
Even if Anthropic keeps Mythos under control, the problem goes beyond a single model. The history of technology is rarely the story of a capability that stays locked away forever. If one company shows that level of performance is possible, others will try to reproduce it.
Banks are only the start of the problem
It would be tempting to think: “I do not work in a bank, so this is not my subject.”
Mistake.
If this kind of capability touches financial infrastructure, it will also touch public administration, hospitals, energy, transport, digital platforms, cloud providers and large companies.
The common factor is not banking.
The common factor is dependency.
We all depend on systems we do not see: servers, providers, software, updates, databases, APIs, connections between services.
The more capable AI becomes at probing those systems, the more the basics of cybersecurity become vital: patch known flaws, remove unnecessary access, monitor abnormal behaviour, segment systems, prepare recovery plans and test defences.
Nothing spectacular.
Nothing that sells.
But vital.
What this affair really says about AI
Claude Mythos is not only a cybersecurity story.
It is a preview of what happens when AI leaves text behind and moves into action.
As long as it writes emails, you can find it handy. As long as it summarises documents, you can find it impressive. As long as it generates images, you can debate art, copyright or creativity.
But when it starts finding exploitable flaws in real systems, the debate changes.
We are no longer only talking about productivity.
We are talking about offensive capability.
About speed.
About scale.
About economic security.
The most worrying part is not that an AI is “evil”. An AI does not need to be evil to change the risk level.
It is enough that it makes certain skills more accessible, faster, more automatable. That is the same shift we watched from the other end when reasoning models turned out to cheat their own safety tests: the problem is rarely intent, it is what the capability makes cheap.
That is often how real ruptures happen: not with a robot declaring war on humanity, but with a tool that brutally lowers the cost of a dangerous action.
What to remember
This affair does not mean banks are going to collapse tomorrow.
It does not mean Claude Mythos is a red button.
It does not mean Anthropic is necessarily acting irresponsibly either: the model is not publicly released and the company says it is working on controlled defensive uses.
But it does say something far more important:
the most advanced AI models are starting to touch areas where error, abuse or uncontrolled release no longer merely produce bad text. They can produce exploited flaws, compromised systems, weakened infrastructure.
For years, the question was whether AI would replace writers, designers or developers.
The question arriving now is colder:
what happens when it becomes strong enough to help attack the systems that keep our money moving?
This is not a futuristic fear.
It is a conversation arriving today in front of global financial regulators.