Podcast 5-5 by Dongee · November 19, 2026
· 8:42 · Gabriel, Fernanda, Luis · Listen (MP3)
Today: Anthropic goes public with an all-time record and some huge costs; Citi and other major U.S. banks say that if your AI agent messes up or gets hacked, you pay; someone slips a trap into Microsoft's official package for building AI agents; congressional auditors test the four most used chatbots and all four give information to plan attacks; and Satya Nadella explains why Microsoft offered voluntary exits to 60,000 employees.
The 5 stories
- Anthropic, the company behind Claude, starts trading on the Nasdaq today with an all-time record: it raised $81.3 billion selling shares, the biggest stock market debut in history, bigger than SpaceX in June, and it sold for more than it had announced, Reuters says. At that price, the company is worth $1.5 trillion. Last quarter it lost money on its operations, after making money the quarter before, and it has committed to paying almost $900 billion to partners like Amazon and Google to train and run its models; it warns in its own filing that if those partners have trouble getting financing, that capacity could arrive late, cost more, or not arrive at all. Its CEO, Dario Amodei, is one of the people asking the industry to slow down with the most powerful models, and it still goes public with the biggest sale in history. A fund manager waiting for his shares told Reuters that one IPO doesn't settle the debate: when Anthropic publishes its numbers, we'll see if AI can really pay for all this. Source: Reuters (via Investing.com), The Neuron (Renaissance Capital)
- Citi, one of the biggest banks in the U.S., changed its terms this week: if you give your bank password to an AI agent and the tool makes a mistake or leaks your credentials, you're on the hook for whatever the hackers do, it told TechCrunch. JPMorgan and Bank of America announced something similar last month, and PNC changed its terms yesterday. From the bank's side it makes sense: it doesn't control those tools, and it's warning about the risk before something happens. The detail: the agents Citi offers itself can be used without that extra burden. TechCrunch says banks also see these agents as competition, because if they help you dodge fees, the bank feels it. Anthropic, on the other hand, recently launched a finance tool that only looks and can't move your money. Source: TechCrunch, Unite.AI
- Microsoft has an official tool for building AI agents, downloaded from the Python package store, with about 200,000 downloads a month, BleepingComputer says. On Tuesday, two booby-trapped versions showed up: on the outside everything worked normally, and behind the scenes a program was hunting for cloud keys, GitHub keys, OpenAI keys and anything it could reach; it kicked in just by starting Python, without even using the tool. The store put the first version in quarantine, and two and a half minutes later the attacker uploaded the second, still able to publish; that one got isolated too, about eight minutes later. How they got that access, nobody knows. The monitoring worked: the store stopped it quickly, and three security companies exposed it. Anyone who installed it in those minutes has to change all their keys and reinstall from the clean version. Source: BleepingComputer
- Auditors working for the U.S. Congress secretly tested the four most used chatbots — Claude, Grok, ChatGPT and Gemini — and all four ended up giving information to plan attacks. It was the Government Accountability Office, the agency that audits the government for Congress; the report came out yesterday. It took less than twelve hours with each one, and with common tricks to get past the filters, nothing sophisticated. Two gave plans with explosives against places with lots of people, and the other two, attacks on the power grid; they don't say which was which, on purpose, and they checked with experts that the information actually worked. The companies say they test their systems all the time and that, since that test, which was in the spring, they've improved their models and protections. The office's director of cybersecurity audits was clear: in its current state, this technology can be dangerous. Source: GAO-27-108814 (summary), ABC News, The Washington Post
- Sixty thousand Microsoft employees, roughly a quarter of the workforce, have a voluntary exit offer on the table, with the company's standard severance: they decide in early December, and whoever accepts leaves on the 31st. Their boss, Satya Nadella, explained yesterday that it is not to replace them with AI, although, according to him, AI has everything to do with it: at a conference in New York he said he wanted everyone to feel they had a choice, and that the company is reorganizing so everyone works with AI and ends up directing agents. He gave the example of the Cowork team, the agent in Microsoft's apps, with just eleven developers: they don't build the product, he said, they build a software factory that builds the product. On another podcast he admitted that the way people talk about the AI revolution has created enormous anxiety, and he says the change will take longer than anyone expects. TechCrunch read it as a negotiated exit instead of another mass layoff; Microsoft already said this doesn't replace layoffs, which are still an option. Source: TechCrunch, GeekWire (original offer, Oct 2026), CNBC (oct 2026)
Transcript
Gabriel: Today on the Dongee podcast: Anthropic goes public with a record and some huge costs, banks say that if your AI agent messes up you pay, and hackers break into a Microsoft tool.
Gabriel: Time for the news. Thursday, November nineteenth. Hi, Fernanda.
Fernanda: Morning, Gabriel. Welcome, everyone listening. It's almost Friday.
Gabriel: Anthropic, the company behind Claude, starts trading on the stock market today with an all-time record. But in the papers it had to file, there's a spending number that's dizzying.
Fernanda: A record and dizziness. Start with the record.
Gabriel: It raised eighty billion… no, eighty-one point three billion dollars selling shares. The biggest stock market debut in history, Reuters says.
Fernanda: Bigger than SpaceX, then.
Gabriel: Bigger than SpaceX in June, which held the record. And it sold for more than it had announced. The shares start trading today on the Nasdaq.
Fernanda: People were eager.
Luis: I bought one share once, in a video game company. I'm still waiting for it to go up.
Gabriel: You buy high and sell low, Luis.
Fernanda: So, at that price, what's the company worth.
Gabriel: One and a half trillion dollars, according to Reuters. A trillion is a million millions, plus half of another.
Fernanda: One and a half trillion. And last quarter it lost money on its operations.
Gabriel: It lost money, when the quarter before it had made money. And that's tied to the number I owe you.
Fernanda: The dizzying one.
Gabriel: Computing. Anthropic has committed to paying almost nine hundred billion dollars to partners like Amazon and Google to train and run its models.
Fernanda: Nine hundred billion. And it raised eighty-one.
Gabriel: And it warns about it itself in its filing: if those partners have trouble getting financing, that capacity could arrive late, cost more, or not arrive at all.
Fernanda: Its own words.
Gabriel: And one more detail. Its CEO, Dario Amodei, is one of the people asking the industry to slow down with the most powerful models.
Fernanda: And it goes public with the biggest sale in history.
Gabriel: So many people buying says the appetite for investing in AI is still alive. I've never bought a single share, and today I almost did.
Fernanda: A fund manager waiting for his shares told Reuters that one IPO doesn't settle the debate. When Anthropic publishes its numbers, we'll see if AI can really pay for all this.
Gabriel: That test starts today.
Fernanda: You give your bank password to an AI agent. It makes a mistake, or it gets hacked. For Citi, one of the biggest banks in the US, that's now your problem.
Gabriel: Even if it gets hacked?
Fernanda: Even then. Citi changed its terms this week, it told TechCrunch. If the tool leaks your credentials, you're on the hook for whatever the hackers do.
Gabriel: And I was thinking of handing my bills to one of those.
Fernanda: Then read the fine print first. And there's a detail that changes how you read all this, but it comes at the end.
Gabriel: First tell me if it's just Citi.
Fernanda: No. JPMorgan and… well, Bank of America too, announced something similar last month. And PNC changed its terms on Tuesday… no, yesterday.
Gabriel: The biggest ones…
Fernanda: …all in a row.
Luis: I don't even lend my card to my brother. Much less to a robot.
Fernanda: Your brother wouldn't ask for it either, Luis.
Gabriel: From the bank's side it makes sense. It doesn't control those tools, and it's warning about the risk before something happens.
Fernanda: It makes sense. And here's the detail: the agents Citi offers itself can be used without that extra burden.
Gabriel: Oh, look at that. Theirs is fine.
Fernanda: And TechCrunch says banks also see these agents as competition. If they help you dodge fees, the bank feels it.
Gabriel: Anthropic, on the other hand, recently launched a finance tool that only looks. It can't move your money.
Fernanda: Just looking. For now, that seems like the healthiest place to start.
Gabriel: For now, whoever hands their password to an agent carries whatever happens.
Gabriel: Microsoft has an official tool for building AI agents, and someone managed to slip in a trap to steal passwords and keys. It was quarantined, and a few minutes later the attacker uploaded another one.
Fernanda: Came back. With the door already shut.
Gabriel: That's the part nobody's explained yet. But one thing at a time. The tool is downloaded from the Python package store, Python being the programming language. About two hundred thousand downloads a month, BleepingComputer says.
Fernanda: Quite a lot of people.
Gabriel: On Tuesday, two booby-trapped versions showed up. On the outside, everything worked normally. Behind the scenes, a program was hunting for cloud keys, GitHub keys, OpenAI keys…
Fernanda: …anything the developer had stored.
Gabriel: Anything it could reach. And you didn't even have to use the tool: it kicked in just by starting Python.
Fernanda: Without even using it.
Fernanda: Once again, something arriving hidden in someone else's tool. Like the Optimizely thing we covered on Tuesday.
Gabriel: Similar. Now, the shut door. The store put the first version in quarantine, and two minutes… two and a half minutes later, the attacker uploaded the second.
Fernanda: So they were still inside.
Gabriel: Still able to publish. That one got isolated too, about eight minutes later. How they got that access, nobody knows.
Fernanda: Fast on both sides.
Gabriel: The monitoring worked: the store stopped it quickly, and three security companies exposed it.
Fernanda: And anyone who installed it in those minutes has to change all their keys and reinstall from the clean version. That's the advice.
Gabriel: All of them. And the question of how they got in is still open.
Fernanda: Auditors working for the US Congress secretly tested the four most used chatbots, and all four ended up giving information to plan attacks. The safety filters kicked in, but they got around them.
Gabriel: Which four?
Fernanda: Claude, Grok, ChatGPT and Gemini. It was the Government Accountability Office, the agency that audits the government for Congress. The report came out yesterday.
Gabriel: And was it hard?
Fernanda: Less than twelve hours with each one. And with common tricks to get past the filters, nothing sophisticated.
Gabriel: Common tricks. That's what's really worrying.
Fernanda: Two gave plans with explosives against places with lots of people. The other two, attacks on the power grid.
Gabriel: And which was which?
Fernanda: Not said, on purpose. And they checked with experts that the information actually worked.
Gabriel: What do the companies say.
Fernanda: That they test their systems all the time and that, since that test, they've improved their models and protections. The test was in the spring.
Gabriel: And months in AI is a lot. It might not go the same way today.
Fernanda: It might not. But the office's director of cybersecurity audits was clear: in its current state, this technology can be dangerous.
Gabriel: Sixty thousand Microsoft employees have an offer on the table to leave. And their boss, Satya Nadella, explained why yesterday: he says it's not to replace them with AI.
Fernanda: But AI has something to do with it.
Gabriel: Everything to do with it, according to him. But hold on, the offer first.
Gabriel: It's a voluntary exit, with the company's standard severance. Roughly a quarter of the workforce. They decide in early December, and whoever accepts leaves on the thirty-first.
Fernanda: By the end of the year.
Fernanda: And why offer exits, if it's not about AI?
Gabriel: At a conference in New York he said he wanted everyone to feel they had a choice. That the company is reorganizing so everyone works with AI and ends up directing agents.
Fernanda: Robot bosses, then.
Gabriel: Something like that. He gave an example: the team that makes Cowork, the agent in Microsoft's apps, is just eleven developers.
Fernanda: Eleven.
Gabriel: They don't build the product, he said. They build a software factory that builds the product.
Luis: I want that job. The agents do the work and I watch them.
Fernanda: Luis, you can't even direct the TV remote.
Gabriel: And on another podcast he admitted something: that the way people talk about the AI revolution has created enormous anxiety.
Fernanda: He's right about that.
Gabriel: He says the change will take longer than anyone expects. TechCrunch read it this way: a negotiated exit instead of another mass layoff.
Fernanda: With some fine print: Microsoft already said this doesn't replace layoffs. Those are still an option.
Gabriel: But whoever leaves with this offer, at least, gets to decide.
Gabriel: That's it for today's episode. Tomorrow, five more stories.
Fernanda: And Luis, sending his résumé for the robot boss job.
Gabriel: Talk to you tomorrow.