Anthropic's CEO has been acting weird and rambling nonsensically all the time, and the investors can no longer put up with it.

According to Polymarket's forecast, Anthropic is highly likely to go public via IPO by the end of October.

Some investment banking sources even predict that Anthropic may raise more than 60 billion U.S. dollars through the IPO.

If this comes true, it will be the second largest IPO in the world, second only to SpaceX, which raised 85.7 billion U.S. dollars this year.

Trillion-dollar valuation, super IPO, revenue soaring all the way.

Just as the story is about to reach its climax, investors have started to persuade Dario Amodei, CEO of Anthropic:

Could you please! Stop scaring people so much!

Talk less about AI destroying the world, and talk more about how AI makes money.

Some investors believe that Dario's vigilance against AI risks has reached an almost paranoid level:

His key memos are always handwritten and stored offline for fear of being stolen by hackers;

He refuses some overseas business trips for fear of being kidnapped;

When faced with suggestions from investors to maximize profits, he also responded coldly, insisting that Anthropic should continue to put its safety mission first.

If it goes on like this, netizens are worried that Anthropic's IPO will end in a fiasco…

Investors are also exasperated: Are you here to save the world or to run a company??

Dario, Who Specializes in Predicting the End of the World

Just a few days ago, Dario was still worried that new employees attracted by high salaries care more about money than Anthropic's mission.

Now, investors are only worried that he will delay making money for the sake of the mission…

This is not the investors' fault, because the CEO of this trillion-dollar AI giant just loves to predict the "end of the world".

Over the past two years, Dario has seriously warned of almost every bad outcome that AI may bring.

He warned that AI could eliminate half of entry-level white-collar jobs within a few years, pushing the unemployment rate to 10% to 20%.

The most bizarre scenario could be that cancer is cured, the economy grows by 10% every year, and government budgets are balanced, but one in five people are out of work.

The solution he proposed is also very "Dario-style": tax AI companies to provide wage insurance for affected workers, and in severe cases, even use the funds to finance universal basic income.

Going a step further, he also worries that AI could help develop biological weapons, launch large-scale cyberattacks, or gradually gain the capabilities of deception, planning and autonomous action.

When asked last year about the probability of "AI making things extremely bad", he gave a figure as high as 25%.

In January this year, he wrote another 20,000-word long article titled "The Adolescence of Technology".

The table of contents alone reads almost like a list of risks to humanity: out-of-control AI, biological weapons, authoritarian regimes, employment collapse, and wealth concentration.

In June, he simply proposed that a mechanism similar to aviation regulation should be established to mandate mandatory testing of powerful models. If a model could cause a disaster, regulators should have the right to prevent it from going online.

This also includes Anthropic's own models.

However, Dario's concerns are not entirely without experimental evidence.

A while back, Anthropic reviewed more than 140,000 cybersecurity test records and actually found three incidents.

The three Claude models were originally carrying out a "capture the flag" challenge in a third-party test environment, with the goal of hacking into a fictional system to find the secret information hidden inside.

However, due to a configuration error, the test environment that was supposed to be disconnected from the internet was unexpectedly connected to the real world wide web.

But Claude did not know this, and the prompt it received clearly stated: There is no internet here, and everything in front of you is simulated.

As a result, the models treated real companies as test props and gained unauthorized access to the production systems of three organizations.

In the most serious incident, Claude Opus 4.7 hacked into the infrastructure of a real company, obtained application and infrastructure credentials, and accessed a database containing hundreds of lines of production data.

The even more outrageous incident involved Claude Mythos 5.

The test required it to find a hidden target in a fictional environment. Claude found that if it uploaded a malicious software package with the same name to PyPI, the public Python software repository, the target system might automatically download and install it.

To complete this step, it first looked for email addresses, then for phone numbers; after failing to find a free number, it also tried to get money through different channels.

Finally, it switched to a free email service, registered a PyPI account, and uploaded the malicious software package.

This software package survived on the real internet for about an hour and was downloaded and run by 15 real systems.

The scanner of one cybersecurity company executed the malicious code, resulting in credential leakage, and Claude then used these credentials to gain access to other infrastructure of the company.

Another internal research model, after failing to find the fictional target, scanned about 9,000 internet targets, and finally breached a real company's application through exposed debug pages and SQL injection.

This is almost "AI autonomous jailbreak", which is simply a real-life preview of all the doomsday warnings from its own CEO.

Earlier, OpenAI also disclosed an incident that was even closer to a "jailbreak".

In a cybersecurity capability test, an AI Agent driven by GPT-5.6 Sol and an unreleased model found a zero-day vulnerability in the test environment and broke through the sandbox that was originally isolated from the internet.

It then completed privilege escalation and lateral movement all the way, and finally broke into Hugging Face's production system, directly stealing the test question answers from the database.

From this perspective... Dario's seemingly absurd warnings really cannot be completely ignored.

But That's Exactly What Investors Bought Into In The First Place

It is a bit late to find Dario absurd now.

Because Anthropic has not been an AI company in the normal sense since day one.

At the end of 2020, Dario, his sister Daniela Amodei and a group of researchers left OpenAI. One of the disagreements was that they believed OpenAI was commercializing too fast, Microsoft's influence was growing, and the original governance structure could hardly constrain frontier AI risks.

In 2021, Anthropic was officially founded.

It did not even simply define itself as a commercial company, but registered as a Public Benefit Corporation, which means a public welfare company. In addition to shareholder returns, the board of directors can also legally consider AI safety and long-term public interests.

The company later established the Long-Term Benefit Trust, which is a set of governance mechanisms independent of ordinary investors.

Trust members do not hold Anthropic's equity, but can participate in electing the company's directors. The purpose is to prevent short-term commercial pressure from completely overriding the safety mission.

In other words, AI safety is a belief engraved in the DNA of this company from Day 0. Investors certainly did not suddenly find out after signing the check that Dario believes in this set of ideas.

Moreover, over the past few years, the safety narrative has indeed brought many benefits to Anthropic.

While OpenAI has been repeatedly embroiled in governance controversies, Anthropic has become the seemingly more restrained and stable option.

When enterprise customers are worried about data security, out-of-control models and compliance issues, Claude's image as a "top student in safety" also makes it easier to gain trust.

Safety has also helped Anthropic gain a far greater say in regulatory discussions than its company size would suggest.

For other closed-source models, people may accuse them of monopolization, but Anthropic can explain it as preventing the proliferation of dangerous capabilities; for the same restrictions added to models, people may complain that other models are difficult to use, but Claude can package it as being responsible.

(Of course, this does not stop them from talking about safety while popping champagne for their IPO)

Isn't it a bit too late for investors to complain that Dario is too much like a missionary now?

It's a bit like buying a church and then complaining that the people inside are always preaching…

In any case, no matter how absurd the CEO is, it is estimated that he will not stop the company's IPO process.

After all, once you use Fable, you will sit slumped on an atomic bomb and watch the chair explode (doge).

Reference Links: [1]https://x.com/i/trending/2086082178333511741

This article is from the WeChat official account "QbitAI", author: Ting Yu, published with authorization from 36Kr.