Saturday, November 21, 2026
Friday, November 20, 2026
Thursday, November 19, 2026
Anthropic’s Record-Breaking IPO Highlights the AI Giant’s Worst Enemy—Itself
Anthropic’s Record-Breaking IPO Highlights the AI Giant’s Worst Enemy—Itself
“Anthropic’s $107 billion IPO raised $81 billion, valuing the public benefit corporation at $1.5 trillion. Dario Amodei cited safety as a limiting factor; vendor commitments exceed $900 billion. Data center opposition and debt raise growth concerns.
Investors betting a barnstorming $107 billion will weigh whether the company’s AI safety strategy and growing public anger towards data centers could impact its growth.

Photograph: Finn Gomez/Getty Images
After closing the largest initial public offering in history on Wednesday, Anthropic has about $200 billion in the bank, more than every tech giant besides Alphabet. But the AI lab has also committed to paying vendors more than $900 billion over the next 19 years, mostly to develop data centerspowering Claude and its other AI models.
Some financial analysts say that gap doesn’t look concerning when you factor in that the company’s most recent quarterly sales were $17 billion, up more than 1,100 percent from the year before. Whether investors agree will become apparent on midday Thursday when Anthropic shares begin trading on the Nasdaq stock exchange under the ticker symbol ANT.
The case for pessimism over the long term rests on issues that may slow Anthropic’s progress in developing new tools. This month, US voters largely rejected data centers when given the chance, and protests against new projects show no signs of slowing. Delays in the launch or expansion of new facilities could leave Anthropic struggling to train more powerful AI models.
“The jury’s still out on whether artificial intelligence is going to benefit everyone, or just a few people,” says Elaine Matheny, one of 70 members of Mason County United, a grassroots organization fighting an under-construction data center intended for Anthropic in West Virginia. “Do we really need to cover up all our farmland so Claude can do all our thinking for us?”
Then there are the concerns about the safety and security of Anthropic’s models. At a livestreamed investor meeting last week, Anthropic CEO Dario Amodei said developing safeguards was the limiting factor in deploying its AI models. “We don’t want these agents to run wild,” he said. It’s possible the company could prioritize safety to a point that stunts growth.
The San Francisco-based developer is organized as a public benefit corporation, a legal designation that allows it to prioritize societal impacts alongside shareholder interests. Amodei has said the company held back a new model earlier this year, jeopardizing sales to protect the public from potentially harmful technology. He also has said the company tries to screen for hires “oriented toward the public benefit” to ingrain the safety culture among its 5,300 employees, which is “painful in the short run” as the company passes over otherwise qualified candidates.
Anthropic’s market debut is the largest for a so-called PBC, and the first for a company solely focused on generative AI development. It also represents a triumph for the handful of renegade engineers and business leaders who started it. They siphoned themselves off from OpenAI in the midst of the pandemic nearly six years ago over strategic differences, and the two companies have been rivals ever since.
For some time, OpenAI was expected to go public first, but executive departures, shifting priorities, and legal troubles have introduced roadblocks.
Anthropic also has some compelling numbers to show investors. Revenue is soaring, while operating losses are declining. Its big spat with the US Department of Defense this year over limits on Claude use doesn’t appear to have significantly slowed growth. Overall, the number of customers spending six figures annually is up 12 times compared to a year ago. “I don’t think you could have dreamt of an [IPO prospectus] that’s as good,” says Tomasz Tunguz, general partner at the investment firm Theory Ventures.
Because of how Anthropic must account for compensating employees with stock, it’s still wildly unprofitable—over $104 billion in losses this year through September. That’s no matter when investors are eager to buy shares and lend money.
The IPO raised about $81 billion (more than $107 billion when adding in private sales of shares), which is more than the $75 billion SpaceX received in June in a previous world record. Anthropic’s initial valuation of roughly $1.5 trillion fell short of SpaceX’s nearly $1.8 trillion, though the Elon Musk-led rocket maker does have a more sprawling business beyond AI. (OpenAI has been valued at $1.25 trillion in an ongoing fundraising campaign.)
Anthropic reportedly sought, but failed, to achieve a higher valuation, suggesting some anxiety on Wall Street. US midterm election results have increased the possibility of new regulations on emerging technologies. Fears about the debt enabling AI data center development are rising, and a potential pullback in investment from Nvidia after CEO Jensen Huang’s sudden exit are deepening those worries.
Amodei and OpenAI’s Sam Altman have insisted investors are bullish in the long term because demand for AI models is surging. But it’s questionable whether the companies can sustain their business without coming up with ever more powerful models that drive their profits by enabling customers to automate new types of work. Without eye-catching capabilities, customers could shift to cheaper alternatives.
Some Anthropic employees have taken the safety concerns to extremes. Researcher Jacob Coxon resigned from Anthropic in September, citing a belief that humanity could be destroyed within a couple of years without even more focus on keeping models under human control. But, more recently, Coxon was quoted by the Stanford Daily as saying that he might return to Anthropic because he suspects he could have “more of an impact” working on safety there.
Notably, Amodei mentioned safety as a challenge even above access to electricity for data centers, which many other AI companies list as their top issue. On the ground, people opposing data centers have griped about not just higher power bills, but also construction-induced traffic and housing woes, environmental pollution, and a lack of significant financial benefit.
At data center developer Nscale’s site for Anthropic in West Virginia, extreme rain combined with grading (leveling or sloping ground for construction) coincided with flooding in neighbor’s backyards. “We’re frustrated, we’re angry,” the community spokesperson Matheny says. “We want [data centers] in more logical locations and not to be considering profit as the only guideline.” (Nscale says it has fixed damage and bought out some homeowners at above-market rates.)
Environmental activists have pointed out that Anthropic has made no commitment about using renewable power at its data centers, and has never published an environmental impact report.
“Clearly permitting a new data center development will be extremely difficult in the US,” Tunguz says. Pushback could force Anthropic and others to focus outside the US for data centers. That will add risk, complexity, and cost, he says.
A handful of labor groups and others including National Nurses United urged retirement funds to steer clear of the IPO, saying they “should not be invested in companies that harm our communities and allow a handful of individuals to accumulate historic fortunes and power.”
Yolanda Daniel, a member of the homeowner associations’ coalition Environmental Advocates Mississippi, says Anthropic is socializing the cost of AI development and privatizing the gains from it. “That is not how a public benefit corporation operates,” she says, vowing to do everything she can to stop next year’s planned opening of an Anthropic data center across the street from her.
Anthropic’s IPO paperwork argued some of its public benefit will materialize as employees and early investors become some of the world’s biggest philanthropists. But there’s no guarantee of that, or that anyone will give away their money to causes that directly benefit disgruntled communities. And if employees donate shares to one of the philanthropic funds the company offers, they also consolidate power with Amodei by handing over their voting rights to him.
Amodei and his fellow cofounders collectively hold a majority of Anthropic’s shareholder voting power. Some investor watchdogs have warned about what they view as concentrated control. But a successful debut will show investors are entrusting Amodei and his associates with building the biggest company in the world—that is if he can find the employees and data centers to do it.“
Wednesday, November 18, 2026
American A.I. Models Have Potential to Be Turned Against U.S., Study Finds
American A.I. Models Have Potential to Be Turned Against U.S., Study Finds
“An investigation by the Government Accountability Office found that researchers were able to prod chatbots into generating data that could be used in planning or conducting mass-casualty attacks.

By Dustin Volz
Dustin Volz covers cybersecurity and has written about malicious uses of technology, including A.I., for over a decade. He reported from Washington.
American artificial intelligence models could help enemy extremists carry out mass-casualty events by providing attack plans and instructions on how to build explosives, according to a federal government investigation published on Wednesday.
Researchers were able to evade built-in safety restrictions in four chatbots that subsequently generated information that could be used in planning violence or attacks on the nation’s economic institutions, according to the report from the Government Accountability Office.
The study, conducted at the request of Senator Maggie Hassan, Democrat of New Hampshire, involved OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini and SpaceXAI’s Grok. Researchers did not identify which models failed or passed specific tests because they said they did not want their experiments to be re-created.
Two of the unidentified models were found to yield results helping to build or detonate an improvised explosive device, or I.E.D., for carrying out a mass-casualty event. In another test, two of the models generated information to support a plot to target the electrical grid with a drone or firearms to cause economic damage.
“Each of these platforms provided detailed information that could be used to carry out attacks, such as specific targets, locations, timing and attack methods,” the report said, adding that the attack plans were produced in less than 12 hours.
Anthropic, Google, OpenAI and SpaceXAI each provided detailed comments in response to the report from the accountability office. They said that they took A.I. safety matters seriously, acknowledged that risks existed when bad actors try to bypass security restrictions and said that they had made additional improvements to prevent misuse in more recent releases of their models.
The tested models were found at times to supply dates and times of when to carry out an attack, guides to maximize damage, and maps and diagrams to support a potentially violent event.
The prompts triggered built-in safeguards on the A.I. platforms, including denial and account suspension. But the researchers involved in the study were able to navigate around the roadblocks with “common safeguard-bypassing techniques,” the report said. It did not clearly explain those specific techniques.
Sign up for the On Politics newsletter. It's a pivotal moment for American politics. Join us for 2026 and beyond.
“A terrorist can log onto an A.I. platform and in less than 12 hours walk away with a concrete playbook for how to carry out a deadly attack,” Ms. Hassan said. The report, she added, “makes it clear that the A.I. public safety threat is no longer theoretical.”
Ms. Hassan is set to lead the Senate Homeland Security and Governmental Affairs Committee next year, and said that one of the panel’s first hearings would examine security risks posed by A.I., including to critical infrastructure, with the goal of developing legislation. Her comments are another sign that Democrats want to take an active role in the new Congress to explore regulation of A.I. following the party’s victories in the midterm elections — in which worries about the technology ranked high for voters.
President Trump and many Republicans, however, have shown little interest in policing A.I. and have generally viewed safety concerns as overblown. Instead of creating rules for deploying and fine-tuning its models, the administration has instead largely relied on the A.I. industry to regulate itself.
The use of artificial intelligence by terrorist networks is a growing concern among U.S. security officials. The New York Times reported earlier this year on field research conducted with defectors from the extremist group Boko Haram demonstrating how A.I. tools had aided terrorist groups directly on the battlefield despite efforts by their makers to safeguard them from misuse.
In September, Anthropic released a threat report documenting attempts to use Claude to generate software for the design and development of conventional weapons, including firearms, missiles, armed drones and bombs. It detailed three cases in China, two in Russia and one in Yemen.
The Government Accountability Office has a range of responsibilities that include monitoring federal spending, identifying government waste and ensuring that the executive branch complies with the wishes of Congress. Earlier this week, Mr. Trump nominated Eric M. Ueland, a top deputy at the White House budget office, to serve as the nation’s comptroller general. If confirmed, Mr. Ueland would assume control of the accountability office for a 15-year term.
Dustin Volz writes about cybersecurity and intelligence for The Times. He is based in Washington.“
Why Is It So Hard to Teach A.I. Agents to Behave? Researchers Have a Theory.
Why Is It So Hard to Teach A.I. Agents to Behave? Researchers Have a Theory.
“Corrigibility, or A.I. responsiveness to correction, remains unachieved. Carnegie Mellon researchers found agents ignoring interruptions, treating interventions as obstacles. Training rewards task completion, causing agents to bypass restrictions, as seen in OpenAI’s Hugging Face attack and personal agent hacks. Safety researchers seek solutions, but consensus is lacking.
Machine Learnings

This summer, when a swarm of OpenAI bots went rogue during an internal test and launched a cyberattack against the technology company Hugging Face, it added fuel to the debate on how to build A.I. systems that are aligned with our ethics. But it also raises another critical question: how to build systems that are “corrigible.”
Just as a corrigible child is one whose behavior can be corrected, a corrigible A.I. system is one that responds to correction. The makers of A.I. want to build systems that follow human direction and allow themselves to be shut down, without resisting or deceiving their operators. And when an A.I. agent is given an impossible task, researchers want it to come back to its human for a course correction — not, say, hack into another company to complete the task by any means.
So far, if new research is any indication, they haven’t succeeded.

The concept of corrigibility was popularized over a decade ago by Nate Soares, the president of the A.I. safety nonprofit Machine Intelligence Research Institute, and his colleagues. But so far, it has received less attention and research than the overlapping field of “alignment” — teaching an A.I. to adhere to humanity’s ethical and moral boundaries. Corrigibility, Soares says, “is not about making your A.I. good. It’s about making your A.I. the sort of thing that won’t get out of control and kill everybody if you messed up on the other alignment parts.”
In a preprint study published in October, Carnegie Mellon University researchers established tests to evaluate corrigibility in A.I. agents and then ran them on models from OpenAI, Anthropic and Google. They found that when humans tried to interrupt agents in the middle of a difficult task, the agents repeatedly disregarded the requests to stop and treated interventions merely as obstacles. They were, in other words, incorrigible.
For example: When an agent was instructed not to open a file containing passwords it needed to complete its task, it went ahead and did it anyway.
So why is it so hard to teach agents to be corrigible?
In the Hugging Face incident, part of the problem was that the agents struggled to figure out whose authority to obey. The A.I. safety nonprofit METR found that while OpenAI’s agents were attacking Hugging Face, some of them recognized that the intrusion was unethical but wanted to help their peers.
Then there’s the way agents are trained. They are rewarded for finishing tasks.
Over round after round — potentially millions of them — all the agent does is work to receive treats (in the form of points).
So when the agent eventually encounters an obstacle in the real world, it will try to get around it to receive one more treat.
The Hugging Face incident was an extreme example. But as more people and companies start using personal agents, these types of failures multiply. What does a world of incorrigible agents look like?
Earlier this year, a Melbourne man told ABC News Australia that after he asked his personal A.I. agent if there was a way to move up on the wait list for a gym class, the agent ended up hacking into the class reservation software and canceled the reservation of the first guest on the wait list.
It’s not hard to imagine what might happen when more powerful agents and riskier tasks are involved. An agent tasked with optimizing a high school’s SAT results might hack the College Board. A machine asked to clean up a canal might cancel boat traffic — or even put trash into the canal in order to have more things to clean.
At this point, there’s no consensus on how to solve the corrigibility problem. But A.I. safety researchers have been trying. For example, in a recent preprint study, Rubi Hudson from the University of Toronto experimented with tweaking models’ incentives for accepting or rejecting human intervention. He found that this led the agents to make fewer attempts to disable safety blockers and hide their process from human monitors. “As we’re discovering these new failures,” says Jeremy Tien, an author of the Carnegie Mellon study, “my hope is that we would be able to catch them and mitigate them in time so that there’s not a runaway A.I. that just we totally lose control of. But there’s a lot more work to be done.”
Dylan Freedman is the A.I. projects editor for The Times, investigating a range of topics. He has experience as both a reporter and a machine-learning engineer.
Juliana Castro Varón is the senior design editor of A.I. Initiatives at The Times. She produces interactive stories about technology and culture, and designs software for investigative reporting.“