Could Mario in Minecraft Cause the Next Financial Crisis?

Hi, I'm Harry. If a friend forwarded this to you, welcome, and you can subscribe here to get it every week. If you enjoy it, I'd be grateful if you passed it on.

Observations and practical guidance from someone deploying AI inside companies every day.

No, seriously. AI has made it cheap to take old software apart, and games were only the start.

Anthropic and Hg are setting up a consultancy to put AI engineers inside European businesses, including Hg's own portfolio. It is the third time this year a frontier lab has teamed up with private equity to do the deployment work itself. Where the labs are putting their effort is a pretty good guide to where AI really is today, and to where your own focus should be.

In the same week, Walmart agreed to let AI shopping agents buy for its customers, even as the big US banks moved to restrict them. Siemens says AI could take factory automation from about a third of the work to 80 per cent.

Elsewhere, OpenAI trimmed its valuation target to $1.25 trillion, and the SEC has put AI on its exam list.

But first, my take on what fans' efforts to take old video games apart can tell us about the software our businesses depend on.

In today's Acquisition Intelligence:

From The Trenches

Video games have just had their unforeseen AI disruption moment.

Everybody expects AI to change industries, and most of us have a picture of how that is supposed to happen. Then little use cases appear that push things in a direction we weren't expecting. In gaming, enthusiasts have started taking existing games apart and rebuilding them into things their publishers never imagined.

Following the September launch of Claude Opus 5.5, gamers have been using increasingly capable coding agents to take old games apart and get them running in places their publishers never intended. Kotaku covered browser versions of Halo and Call of Duty: Black Ops. Others have gone further and started splicing games together.

There is Minecraft inside Elden Ring, Minecraft inside Super Mario 64, and Minecraft inside GTA V running on an iPhone. Apparently, having one game to play at a time was becoming restrictive.

I find these small, obsessive pockets of the internet useful to watch. People have a reason to push the technology, plenty of patience and very little concern for whether the result fits an industry forecast. Sometimes their idea of a good weekend gives the rest of us an early look at a much bigger change.

What caught my attention was the work behind the clips. People were using AI to help understand and change software they hadn't written, sometimes without its original source code. A game is a convenient place to experiment with that ability, but the security news this week showed how much further it could reach.

A week of warnings

On October 19, Reuters reported that one hacker had used AI agents to break into websites belonging to more than 100 companies.

Two days later, Wiz disclosed AI-assisted discoveries of serious flaws in PHP, software used to run a large part of the web. Some weaknesses had been present for more than a decade. Researchers had used AI to help turn difficult technical problems into working attacks, prompting urgent fixes.

The same week also showed how exposure can spread through the services developers depend on. Hosting provider Railway warned that customer credentials may have been exposed, and attackers put credential-stealing software into a release of LiteLLM, a tool developers use to connect to AI models. In both cases, the concern extended to the other systems those credentials could unlock.

Warnings about AI and cybersecurity have been coming for a while. Anyone following Dario Amodei will recognise the concern: Anthropic restricted access to Mythos Preview in April because of its cybersecurity capabilities, and Amodei has since argued for slowing the advance of the most capable models to give safeguards time to catch up.

But this week felt like a watershed moment, capped by Anthropic's deputy security chief Jason Clinton posting on October 22 that things would get worse before they got better. He urged security leaders to ask their boards for help with the basics. Coming after a week like that, I wanted to understand what those boards should actually be worried about.

What is actually exposed?

With cybersecurity becoming a more immediate concern, I started digging into how these attacks work and which businesses might be most exposed. We're used to hearing about stolen accounts and companies being held to ransom. What does AI change about the way someone gets in, and how much of the business can they reach once they do?

The games offer a useful starting point: the person taking one apart has the program on their own machine. Decompilation reconstructs readable code from that finished program, and AI helps with the work of understanding it. The fan can experiment without needing access to Nintendo's systems.

Some business software is available in the same way. A company might download and install a database such as Microsoft SQL Server, and someone looking for weaknesses can obtain their own version to study. They get the program, without the company's records, but a flaw they find could work against other exposed installations of the same product.

That is the connection to this week's PHP news. The software is freely available for researchers to inspect, yet a weakness in it can leave many unrelated websites needing a fix. An attacker may only need to solve the problem once before looking for other businesses that haven't updated.

That made me wonder about the software most of us actually spend our day using: Salesforce, Gmail and Outlook. We haven't downloaded those services, so are our emails and customer records exposed in the same way?

The provider runs and updates the underlying service. You manage access to your company's account, and an attacker may get in by stealing a login, tricking an employee into approving access or compromising an application connected to it. Your emails could be stolen while Outlook itself keeps working perfectly.

That puts some very familiar precautions back in the picture: unique passwords, multifactor authentication, ideally with a passkey or security key, and care over which applications can access company information. Providers can still have software vulnerabilities, but an attacker who obtains valid account access may never need to find one.

The financial consequences can arrive long before a company loses the ability to operate. After Yahoo disclosed major data breaches during its sale process, Verizon cut the agreed price for its operating business by $350 million. Stolen records had become a problem at the negotiating table.

What happens when someone gets in?

The damage depends on what that access allows. An employee's email might be used to impersonate them and request a payment; customer records could expose people to fraud and leave the company facing claims. The stakes rise further if the attacker reaches the systems staff need to do their jobs.

Ransomware can lock files and prompt a demand for payment, often alongside threats to publish stolen data. Attackers can also alter records, making it difficult to trust what the system says.

Management may then take systems offline to contain the incident. A business can end up unable to operate even when the attacker hasn't destroyed the software it needs.

Clorox gives a fairly tangible example. After its August 2023 attack, the company took systems offline and resorted to manual ordering and processing at reduced capacity. It still had cleaning products to sell, but its ability to get them to customers was impaired.

On October 4, it warned that quarterly sales would fall 23–28% from a year earlier, largely because of the disruption. The following day, Reuters reported its shares down about 8%. Investors were having to put a price on how much business the company could lose while it recovered.

For a smaller company, I'd translate that into cash. How much can it sell if staff cannot see inventory or issue invoices, and how long can it keep paying wages while receipts fall? Restoring a backup helps, but someone still has to establish that the systems are safe and reconcile the work done during the outage.

That is why I'd worry most about businesses that struggle to update essential software and cannot operate for long without it. A manufacturer may need to stop production to make a change, while a healthcare provider still has patients to treat during an outage. The useful question is which part of the business stops when a particular application disappears.

How worried should we be?

Once several businesses depend on the same software or provider, it becomes easier to see how three isolated incidents could become ten, twenty or a thousand. AI can help attackers investigate more targets, but they don't necessarily need a separate break-in at every company to disrupt them all.

Individual attacks are already a regular occurrence. The FBI received 3,611 ransomware complaints in 2025, up from 3,156 in 2024, with other incidents going unreported. Those figures don't give us the odds for a particular company, but criminals already have a business in finding access and demanding payment; AI gives them a way to investigate more targets with the same resources.

Shared software creates two ways for that pressure to spread. A flaw can be exploited at several companies running the same product, as the PHP example illustrates. Or an attack on one provider can leave many customers without an essential service.

The 2024 CDK attack disrupted software used by more than 15,000 retail locations across North America, sending car dealers back to manual paperwork. Those dealers didn't each have to suffer a separate break-in to lose access to the systems they needed.

Change Healthcare's outage that year interrupted insurance claims and payments, leaving healthcare providers struggling to collect cash. By its second-quarter results, UnitedHealth had provided more than $9 billion in advances and interest-free loans to support providers. One supplier's problem had become a financing problem for its customers.

Could a disruption reach the financial system itself? The New York Fed has modelled how interrupted payments at a major bank could leave other institutions short of expected funds. But a website breach doesn't automatically grant access to a trading system, and a wider crisis would require disruption to overwhelm the controls and recovery arrangements intended to contain it.

That leaves me most concerned about the connections between ordinary businesses and the software they depend on. The gamers are demonstrating how much easier it is becoming to investigate a program. The business consequences depend on what that program controls and how many companies need it to keep working.

What I'd ask management

I'd bring that back to the systems we need to take orders, serve customers and collect cash. Which do we maintain ourselves, and which do we rely on somebody else to run? That tells us where to ask about overdue fixes, account access and the supplier's recovery arrangements.

Then I'd walk through two weeks without the most important one. How would staff keep working, what could we restore, and how much cash would we need while receipts were delayed? A recovery plan is more useful when someone has demonstrated that it works.

There is an encouraging side to the gaming example too: the same tools can help defenders understand and repair software that was previously expensive to investigate. But some fixes already have a clear answer and are waiting for budget or permission to interrupt operations. I'd want to know which of those decisions were sitting with us.

At any rate, it's a good time to revisit the software we rely on and hold an internal security review, because it's going to be a bumpy ride from here.

From The Builders

Nassim Nicholas Taleb on white-collar work (X). Taleb says every white-collar profession is about to be decimated by AI, and points to this month's cuts: 3,500 at Amex, 5,000 at Barclays and about 9 per cent of eBay. Expertise that used to be expensive is getting cheap. I'd look at which of your cost lines are mostly people reading, checking and reconciling.

Bryan Chan on a model finding a hidden message (X). One Fable hid a secret message in an ordinary-looking photo. A fresh Fable, asked only whether anything was hidden, read it back word for word. Models are now very good at reading things nobody meant them to read.

Social Capital on where the AI money goes (X, October 24). Of the roughly $2.7 trillion going into AI this year, Gartner puts about $28 billion, around 1 per cent, on the models themselves. Services ($577 billion) get more than software ($462 billion). That's why the labs are moving into services, as Anthropic did with Hg this week.

News Digest

Anthropic and Hg Build a Consultancy for Europe

Anthropic and Hg, the London software buyout firm, have agreed to set up an independent company that puts AI consultants and engineers inside European businesses, the FT reported on October 23. It will work with Hg's portfolio companies as well as other clients, with hubs in London and Berlin and a launch expected by the end of the year. Anthropic is putting in balance-sheet capital and Hg is investing from its funds, including HgCapital Trust. The FT says the aim is a company worth more than $10 billion within five years.

The details:

Why it matters: The labs are moving into work the big consultancies and IT services firms have been selling, and they are doing it through private equity.

My take: The labs have worked out that a model doesn't get adopted on its own. Someone has to sit inside the business and understand how the work is actually done before they can redesign it. That's the work we do at DealSage for mid-market portfolio companies. The largest sponsors now have a lab-backed team to call on. I'd also ask who owns the firm doing the work, because an adviser part-owned by a model vendor will tend to recommend that vendor's model.

Read more

Walmart Opens the Door to Shopping Agents

Walmart said on October 21 that it will let AI agents shop and pay for its customers. Its new Agent Gateway lets approved assistants, including ChatGPT, Gemini and Meta's Muse, search Walmart and Sam's Club stock and buy on a customer's behalf without seeing their personal data or purchase history. From next month, Agent Profiles will let shoppers approve each agent, set spending limits or sign off every purchase, and switch an agent off at any time.

The details:

Why it matters: The largest retailer in the world is treating agents as a sales channel in the same week the biggest banks are treating them as a fraud risk.

My take: A few weeks ago I wrote about what consumer agents could do to business models, and this is the first big company to build a way to serve them directly. Each agent gets its own identity and spending limit, and can be switched off, so nobody has to hand over their password. Every business that sells online will have to decide whether to serve agents, block them or let them scrape the website, and somebody in management should make that decision on purpose.

Read more

Siemens Puts a Number on the AI Factory

Cedrik Neike, the Siemens board member who runs its industrial software and automation business, told Reuters on October 19 that AI could raise factory automation to as much as 80 per cent. Siemens estimates 30 to 40 per cent of factory work is automated today. Neike said manufacturers in high-wage countries like Germany need productivity gains of 7 to 9 per cent a year to stay competitive.

The details:

Why it matters: Operating companies outside finance are now telling investors how much they expect AI to save or earn.

My take: I find these operating numbers more useful than another round of bank job cuts. If you own a manufacturer or a logistics business, you should know what productivity gain the plan assumes and whether anyone has tested it on the line. Siemens tests every change in simulation before it touches the floor, and I'd want a portfolio company to do the same.

Read more

DealSage in the news

Ben Amor, our co-founder and CTO, has been out talking about what actually makes AI work inside a business. He's done two interviews:

The Drawdown: "5 minutes with… Benjamin Amor, DealSage"

Ben on why the biggest AI opportunity for GPs is the CRM nobody likes keeping up to date, and why most of the work at the portfolio company level "is not really AI at all". It's about finally getting the data in one place.

Read it on The Drawdown

Media and the Machine, episode 28: "Ex-Palantir Vet: Your AI Problem Is Really a Data Problem" (30 min)

He draws on a decade of deploying AI at Palantir to explain forward-deployed engineering and why so few pilots make it to production.

Listen to the episode or read the write-up

More coverage is on our press page.

Other Interesting Things I've Read or Seen This Week

OpenAI trimmed its valuation target to $1.25 trillion (FT, October 22). It had pitched $1.4 trillion earlier in the month for a round of up to $30 billion, and some investors pushed back on the price. Some are reportedly waiting to see how Anthropic's IPO is priced. Both labs are working hard this month to demonstrate revenue and defend their valuations.

The SEC put AI and private funds on its exam list (Reuters, October 23). AI, private funds and private credit are among the regulator's exam priorities for fiscal 2027. I'd check that what your firm says about its AI matches what it does.

Dayforce is buying Ribbon AI (Dayforce, October 23). The HR software company is adding AI agents that run parts of recruiting. It's another software incumbent buying the agent instead of building it.

Wipro's large-deal bookings hit a 19-quarter low (Moneycontrol, October 16). The chief executive expects a rebound next quarter. If AI makes old code cheaper to read and rewrite, the firms paid by the hour to do it feel that first.

PE healthcare dealmakers put AI and policy at the top of their worries (WSJ, October 23). Attendees at McDermott's New York healthcare private equity conference said buyers now want proof that businesses can hold up under pressure and are protected from government changes. Buyers in what used to be the safe sector now have two more things to worry about.

Know someone who'd find this useful? Forward this email, or send them your share link. Most new readers come from a friend passing it on.

Acquisition Intelligence is a weekly newsletter on how mid-market companies put AI to work, for operators and the investors who back them.

Questions, feedback, or something you're seeing? Just reply.

P.S. I'm Harry, co-founder of DealSage. Before we build anything for a portfolio company, we start by reading what it already runs, including the old systems and data nobody fully understands any more. If you'd like to see what that looks like for one of your companies, reply and I'll show you.