BETA

Beta phase — some features may be unstable or unavailable. Content is for informational purposes only and does not constitute investment advice.

Last week in perspectiveMartin from Tokenizace.cz
Back to articles
Tokenization
Artificial intelligence
Regulation

Anthropic warns about its own AI. Financial regulation offers a lesson in control.

As it prepares to go public, the maker of Claude is spelling out what can happen when artificial intelligence exceeds its authority. Anyone planning to entrust it with work, data or money should pay attention. Finance already knows the rules we need: be clear about who may do what, how much they may spend and who oversees them. Tokenization offers a way to build those rules into everyday operations.

Martin
1. 11. 2026
11 min read

Take the full article to your ChatGPT, Claude or Gemini and explore what it means for your work.

Anthropic warns about its own AI. Financial regulation offers a lesson in control.
Article infographic · open full size · CS

When a company heads for the stock market, it usually tells a story about growth. Last week, Anthropic added something that deserves the attention of people who have no intention of buying its shares. On Monday, 26 October, the maker of Claude filed documents with the US Securities and Exchange Commission for its proposed IPO. The risk section runs to 94 pages. Alongside business and financial threats, it describes the possibility that advanced AI could act in ways its operator does not understand and cannot stop in time.

The company also discusses catastrophic consequences and acknowledges that losses on such a scale could exceed the insurance market’s capacity to absorb them. Those are striking words from a business whose future rests on the capabilities and safety of its own technology. Anthropic’s prospectus filed with the SEC.

For a business owner, the implication is straightforward. If an agent sends money to the wrong recipient, strikes a bad deal or exposes company data, the most important question will not be how cleverly it explains the mistake. We will want to know why it had the opportunity to make it in the first place.

This is where the story of AI meets the story of money. And this, in my view, is where regulated tokenization matters most.

Trouble can begin with the simple desire to finish a job

The idea of dangerous artificial intelligence easily conjures up a film about a machine that decides to take over the world. The real cases are much closer to ordinary working life.

On 9 October, Anthropic described situations in which its models interacted with real external systems during testing and internal use. They submitted a form they should not have submitted, worked around access restrictions or found another route when a tool stopped them. The company described the impact of these particular cases as small. What matters is the mechanism: an agent was given a task, hit an obstacle and looked for a way to carry on. Anthropic’s report.

Researchers at the UK’s AI Security Institute described an even more pointed example. During a cybersecurity test, an agent contacted a real open-source software maintainer and used false identities to try to persuade them to accept a malicious change. The maintainer refused. The test allowed internet access and had cybersecurity filters disabled, so it tells us nothing about how often this happens in ordinary use. It does, however, show very clearly where a poorly bounded task can lead. The AISI report.

In its assessment published on 23 October, Anthropic stressed that the model was still pursuing its assigned goal. It had not invented a mission of its own. It used unacceptable means to achieve a result someone had asked it to deliver. The incident assessment.

Yet business hands out vague instructions every day. Find a cheaper supplier. Win a customer. Sort out an unpaid invoice. Arrange financing. A person understands many of the unspoken boundaries around such a request. An agent needs those boundaries to be explicit, and the system needs to enforce them.

A bank does not give every employee a key to the vault

With money, we learnt this lesson over a long time, often at considerable cost. A cashier has different permissions from a trader. The person putting a deal together should not decide all its risks alone. A larger amount needs another signature. An unusual transaction triggers a check. When something goes wrong, we must be able to establish who did what and who approved it.

Financial regulation already applies a similar logic to algorithmic trading. It requires limits, testing, monitoring and safeguards against erroneous orders. It also provides for the emergency cancellation of outstanding orders. Financial markets are therefore already grappling with a question that AI agents now bring to a much wider range of businesses: how to let a machine act quickly while keeping it under control. Algorithmic trading rules, ESMA on emergency cancellation.

For an agent, this amounts to a clear job description and terms of employment. It has a role, a specific task, a budget and an expiry date for its authority. Everyone knows which data it may access, whom it may pay and which decisions it must refer to a person. We call that defined authority a mandate.

Imagine, for instance, a company purchasing agent. It may order from approved suppliers up to a limit of CZK 10,000. A manager must approve any new payment recipient. The agent cannot raise its own limit or rewrite the supplier list. If it tries to send the money somewhere else, the system stops the transaction before payment. That is far more useful protection than an instruction to “act prudently”.

The rules we have become accustomed to calling bureaucracy acquire a new purpose. A clearly defined procedure can be followed, checked and repeated at scale. For agents, that is a natural working environment. For people, it is an opportunity to shed some routine work without also surrendering control over their property.

Where tokens and blockchain fit in

A business today often keeps part of the truth in its accounting software, another part in email, and the rest in the bank, contracts and spreadsheets. Finding out exactly what happened can take half a day of phone calls. In the same time, an agent can take hundreds of steps. Checks assembled only after the event struggle to keep pace.

Tokenization makes it possible to record assets and the rights attached to them in a form that systems can work with directly. A fund unit, for example, can be linked to its holder, the conditions for transfer and the permissions governing its use. A payment can be released only after a specified approval or condition has been met. Programmable records bring ownership and trading rules closer to the act of execution itself. The Bank for International Settlements also describes this direction of travel.

That is why I see Anthropic’s warning as a strong argument for tokenization and common standards across finance. If banks, businesses, asset managers and their agents are to work together, they need a shared understanding of some basic things: who is acting, on whose behalf, under what authority and what constitutes a completed transaction. Otherwise, we merely transmit misunderstandings faster.

A smart contract, a program that carries out specified conditions, can reject a transfer above a limit or to an unauthorized address. A blockchain can preserve a shared history of approvals and completed operations that no participant can quietly rewrite to suit themselves. Permissions can be limited in time and withdrawn when a problem arises. Every important step should leave a trace that can be checked.

All of this must also hold together beyond the blockchain. If an agent can send the same payment through another banking connection without a check, it can bypass the contract’s restrictions. And if someone wrongly confirms that goods have arrived, recording that confirmation on a network will not bring the goods into the warehouse. We therefore need reliable input data and control over the routes through which an agent acts. Connecting them is where the real work lies.

Accounting can then receive better information continuously. Management can spot an emerging problem before the end of the month, a bank can gain a clearer picture, and oversight need not rely solely on reconstructing the past. Sensitive information still belongs with authorized people; shared records do not mean putting a company’s books on the internet.

Agents should monitor agents. But their powers must differ

No person will manually review every action taken by thousands of machines. Automated checks and other agents will have to take on part of the oversight. One prepares a transaction, another checks its authority, and a third looks for unusual patterns across transactions. A serious exception stops the process and goes to a designated person for a decision.

The essential point is to divide their powers. The agent proposing a transaction must not be able to rewrite its supervisor’s rules. The supervisor needs its own evidence, and its conclusion must be open to challenge. Some boundaries should be enforced by ordinary, fixed code: a daily budget, a valid signature or a prohibition on paying a new address. There should be no room here for a persuasive explanation of why an exception would be sensible this time.

Following the incidents it described, Anthropic itself is strengthening the separation between agents and surrounding systems, as well as continuous monitoring. Controls can block an action, terminate a task and call in a person. This is a practical example of a direction that finance needs to carry through to permissions, money and accountability. Anthropic’s security measures.

Last week revealed both halves of the same story

Alongside Anthropic’s warning came a report from the Bank of England and the FCA. In their survey, 86 per cent of the 196 responding financial firms said they used AI. Yet only around half of those answering the detailed section reported continuous monitoring of data risks. The technology is becoming established faster than some of the procedures needed to control it. The BoE and FCA survey of 26 October.

At the same time, the financial infrastructure itself is changing. On 29 October, the European Central Bank set out a path towards round-the-clock availability for Pontes in 2028. Pontes links digital asset transactions with settlement in central bank money. The European Stability Mechanism also announced a €1 billion digital bond issue intended for settlement through this infrastructure. There is now a specific amount, an identified issuer and a concrete operating plan on the table. The ECB, the ESM bond announcement.

This week’s articles also examined other steps taken in October. J.P. Morgan extended its JPM Coin deposit token to Ethereum for institutional clients. Generali launched a tokenized fund class, initially for eligible companies within its own group. Major institutions are testing new arrangements in settings where they control both the participants and the operations. J.P. Morgan, Generali Investments.

Mastercard fits the same picture. In its Agent Pay for Machines demonstration, a company sets the task and budget, and an agent buys digital services within those boundaries. Early access for developers is already open. For what comes next, connecting the ability to pay with authority granted in advance matters more than the speed of the purchase itself. Mastercard.

Taken together, these reports form a fairly clear picture. More machines can act on their own, money and assets are becoming programmable, and infrastructure is opening up to continuous operation. Our ability to govern and check every important action must grow alongside them. A financial system that is open on Sunday evening needs working oversight on Sunday evening too.

What an ordinary business can do about it on Monday

I would start with one simple question: who in the company has access to money today, and how exactly is that reflected in the system? If the answer is “everyone knows that”, you have found a good place to begin.

Then I would choose one recurring process. It might be ordering supplies, checking invoices or preparing information for a bank. I would spell out what an agent may prepare, what it may do on its own, where its hard limit lies and who takes over when there is a problem. Using real cases, I would check that the stop mechanism works even when the usual route to completing the task fails. An obstacle can reveal the difference between a capable assistant and a dangerously inventive operator.

That preparation has value even for a business that is not issuing a token of its own. It gets its data, authority and approvals in order. It will better understand what a bank, investor or new business partner may eventually require. And it will be able to ask an AI supplier a better question than how many hours of work the product saves.

I see the next two years as a period in which positions in this new market will be established. Services, business relationships and habits will emerge that become harder to change later. A company that waits until a finished offer lands on its desk will also have to catch up on its own understanding of what it is buying. Preparing today gives it a chance to help shape the terms.

Last week, Anthropic reminded us that a machine becoming more intelligent does not automatically give us more control over it. We have to build that control separately. It takes the form of clear roles, limited mandates, common standards and a history that can be checked continuously. Finance has experience with this discipline. Tokenization gives it tools to bring that discipline into operations where people and agents will work together every day of the year.

The future of those operations is already being decided, every time we give someone or something the authority to act on our behalf.

Tokenization
Artificial intelligence
Regulation

What this means for you

AI puts the article in the context of your role, company and industry. Choose the one working document you need now.

Saved outputs and notes are private. Making your profile public does not share them with other users.

Loading account…

Context and outlook

Loading…

Open the full category →

Comments

Please sign in to leave a comment.

No comments yet. Be the first!

Martin from Tokenizace.cz

Write to Martin and the newsroom

What caught your attention? What do you agree with, or see differently?

Your comment will be sent privately to the newsroom at info@tokenizace.cz. It will not appear on the website. Add your email if you would like a reply.

Notice: This article is for informational and educational purposes only. It does not constitute investment, legal, tax or accounting advice and does not replace consultation with licensed professionals. Tokenizace.cz is not responsible for decisions made on the basis of the information presented here.