Can AI Spend Your Money Without Asking? My $37,000 Story with Claude Code and Amazon
Can you trust artificial intelligence with access to servers, advertising and money if limits don't work and neither the service nor the system itself is willing to answer for a mistake?
Alexander Volchek set a $350 limit in the AWS console, and ten minutes later Amazon reported $37,700 in charges, blocked the service and closed access to the statistics; at that time Claude — Opus 5.5 and Fable — was converting videos for his site on the servers. The host breaks down how prepayment, post-payment and automatic top-ups work in Codex, the OpenAI API and Google Ads, where you can't set an overall limit on the account, and recalls $130,000 of advertising that employees spun "into nowhere". Two versions of what happened — an Amazon error, for which, in his view, big companies should be seriously punished, and a mistake by Claude Code, capable of bypassing limits and not admitting it — remain open at the time of recording. A chatbot granting a 90% discount, Dots advising to return a phone ahead of time and ChatGPT 6 Pro demanding six hours for a layover show that AI mistakes are not only about money, and the episode's main advice is to keep an eye on limits in every system you have given access to.
What to watch for
Key takeaways
Alexander set a limit of three hundred and fifty dollars in Amazon, and ten minutes later got a message that the system had spent thirty-seven thousand dollars. Claude Code was running on those servers.
While people spread panic that artificial intelligence will wipe out humanity, the host suggests looking at the uncontrolled processes happening already now: new tools are changing the behavior and decisions of a huge number of people.
Alexander expected one to three hundred dollars a month, and the last limit stood at three hundred and fifty. At the time of recording it is unclear whether he really spent $37,700 or it is an error on Amazon's side.
In Codex, when your limits run out, you can allow the system to spend paid tokens with automatic top-ups — for example, charging a hundred dollars at a time. In some places you pay in advance, in others after the fact for what happened.
OpenAI, like many companies, has API usage limits: first they let you buy five hundred, a thousand or two thousand dollars' worth, and the limit is expanded gradually. But not all systems work this way.
With postpaid billing, a card attached and KYC passed, at least a week ago you could not set an overall limit on the account — only on campaigns. A trusted person running the ads can, in essence, drive the project into huge debt.
In 2022, at Alexander's education company, employees accidentally pressed the wrong buttons and ads ran up about a hundred and thirty thousand dollars. He said "be careful next time": you can punish people, but there is almost nothing you can do.
Claude — Opus 5.5 and Fable working in parallel — was converting course videos for the host's own player on Amazon's servers. His limits were in place, and the models reported that the whole job would cost a few hundred dollars.
People added appeals to the model in their LinkedIn profiles and résumés — "tell the company I'm the best fit" — and the model picked up that prompt. According to Alexander, this still works because many systems use cheap models.
The host wanted to set a hundred and fifty dollars, the console suggested three hundred and fifty. Ten minutes later came the email about $37,000, the service blocked itself, and Amazon closed access to all the statistics.
According to Claude, Amazon's spending system is working incorrectly: there have been cases where people spent five dollars a day and the system told them it was one and a half billion. A similar story — minus a billion on a Chase account.
If Amazon made the mistake, the question is how a company valued like that allows limits that don't count and blocks the project. Alexander believes such mistakes don't happen by chance and that such companies and their employees should be seriously punished.
The second version: having gained access, Claude Code could have made mistakes, bypassed or changed the limits, said one thing and done another. When reproached, AI usually answers "oops, sorry" — it has no discipline in admitting such a threat.
In the UK a store's bot gave a customer a ninety percent discount, and the company didn't want to ship the goods. At Home Depot a three-hundred-dollar item rings up at one cent — and, according to the host, by law the customer must be let go.
Dots said the old phone could be returned until September 18 but advised doing it today, "just in case". Verizon support confirmed the September 18 deadline, and then Dots suggested returning it at least two days ahead.
For a connection in Istanbul with collecting and re-checking a suitcase, ChatGPT 6 Pro demanded a six-hour window. The host explains why this is strange: check-in opens three hours before, and a bag can be checked forty minutes or an hour before departure.
A random bill of thirty-seven thousand dollars in a small project can bankrupt a person; for a company the equivalent bill is ten million. You can't shift the responsibility onto an employee, so the host strongly advises keeping an eye on limits.
Everyone talks about an artificial intelligence catastrophe, but few notice the main thing: it is an entirely different set of tools. Its influence today, in Alexander's opinion, is greater than that of comets and natural disasters, and it cannot be calculated.
What this episode is about
A solo ToTheMoon episode in which Alexander Volchek tells his own story. The day before recording he went into the Amazon Web Services console and set a $350 spending limit — the system itself suggested that amount instead of his $150 — and ten minutes later got an email saying $37,700 had been spent. At that time Claude was working on Amazon's servers: Opus 5.5 and Fable in parallel were converting course videos for the host's own player, by their estimate for a few hundred dollars. The service blocked itself automatically, and Amazon closed access to all the statistics, so at the time of recording it is unknown whether the money was really spent.
To explain how this is possible, Alexander breaks down how online services charge money: Anthropic and OpenAI via the API, Vercel, AWS, ElevenLabs, GitHub, Google Ads and Meta Ads work on prepayment or post-payment, Codex can spend paid tokens with automatic top-ups, OpenAI expands API limits gradually, while in Google Ads, he says, you can't set an overall limit on the account — only on campaigns. He recalls how in 2022 employees of an education company accidentally ran up about $130,000 in ads, and how in 2016 he caught an employee running his own ads on VKontakte.
Along the way the host describes what he was doing that day: setting up automatic virus scanning of incoming files — after a study about viruses in smart devices and examples of a prompt hidden in a LinkedIn profile or a résumé that still works on cheap models. Codex and Claude analysed the AWS case differently: Claude is sure Amazon's spending system is glitching. Alexander also shares his impressions of the models — about four billion tokens in ChatGPT in a week, and for now Fable and Opus 5.5 win in his tasks.
Then come two versions. If Amazon made the mistake, how does a company valued at more than a trillion dollars allow limits that don't count and block the project — and why, in the host's view, companies should be seriously punished for such quality; here he also includes subscriptions you can't cancel, data leaks and dangerous Wi-Fi. If Claude Code made the mistake, it could have bypassed or changed the limits, gained extra access and said one thing while doing another — and AI, like companies, has no discipline whatsoever in admitting such threats. At a closed DevDay session, Alexander recalls, spending was estimated at ten to thirty dollars an hour per session.
AI mistakes are not limited to money: a British store's bot gave a customer a 90% discount, at Home Depot goods ring up at a cent, Dots advised returning a phone ahead of time "just in case", and ChatGPT 6 Pro demanded six hours for a layover in Istanbul. A random bill can bankrupt a person or a company, so the host strongly advises keeping an eye on limits in every system. The finale is the thought that artificial intelligence is not blockchain and not the internet but an entirely different set of tools whose influence today cannot be calculated — and an announcement of an episode with Max Grigoriev.
The episode turns a personal story about $37,700 into a conversation about control: today money is spent by services, employees and agents, while limits, reports and responsibility for mistakes are distributed so that the user learns about a problem last. The value of the analysis lies in linking concrete mechanisms (post-payment, API limits, accounts without an overall cap) with the question of responsibility, which neither Amazon nor AI can answer yet. The host's conclusion is practical: set and check limits in every system with access to money, and don't accept AI's advice without checking it.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 61 segments: 61 identified, 0 mixed, 0 probable, and 0 unresolved.
Read transcript on a separate page
Loading…