• Why Palantir CEO Alex Karp is wrong about the frontier AI labs.
• Autonomous ransomware is here, a cybersecurity firm claims.
• China considers restricting foreign access to leading AI models.
• Anthropic finds part of LLMs functions like an aspect of human consciousness.
• AI safety standards are slipping, report says.

This week, we have some exciting news right here at Fortune. We’re launching a brand new vodcast called Fortune AI Weekly, which I’m co-hosting with Bea Nolan. You can think of it a bit as an extension of what we do here at Eye on AI—bringing you our thoughts on the biggest AI news of the week, highlighting some of Fortune’s great AI reporting, and sometimes bringing you exclusive interviews with key AI builders, thinkers, founders, funders, and leaders. You can check out the vod on our YouTube channel here.

***

Among the AI news that made headlines last week was Alex Karp’s rant against the foundation model companies. The Palantir CEO went on CNBC Wednesday ostensibly to discuss a new partnership between Palantir and Nvidia to provide a “sovereign AI infrastructure” to the U.S. government and critical industries. The collaboration involves the use of Nvidia’s Nemotron open source models along with Palantir’s Artificial Intelligence Platform (AIP) that is an application layer connecting those models to data, along with offering data security and governance. But that’s not what wound up making headlines. Rather, after prefacing his remarks by saying “I’m not throwing shade” at OpenAI and Anthropic, Karp proceeded to toss Mordor levels of shadow at the frontier AI labs.

“Something has gone completely wrong,” he said. “The basic view among enterprises in this country is ‘I’m going to chillax and waste my time with tokens, I’m going to get no value, and their going to get my IP.'” He then said this was not shade but “reporting.” He doubled down on these points several times, saying that companies were getting no value from the tokens they are purchasing from the frontier labs and that they are risking transferring their crucial business IP to these AI vendors.

So does Karp have a point? Well, kind of. But only if you squint. And much of what Karp said was either self-serving, inaccurate, or contradictory—or all three.

ROI is lagging, but no one is ‘chillaxing’ about it

It’s true that many large companies are worried that they aren’t yet seeing enough of a return on investment from deploying AI and are fretting about how much tokens are costing them, particularly when using the most advanced AI models in agentic use cases. (The fact that many large companies are concerned about this contradicts Karp’s claim that they are simply “chillaxing.”) But certainly some companies are reporting value—particularly in software development and customer service. And, for those that are not, it is often because they have not prioritized the most strategically essential use cases or figured out how to reengineer their workflows across the company to take best advantage of the technology.

At one point in the CNBC interview, Karp said “why are they charging for tokens, if it is so valuable?” He suggested that if the foundation models worked as well as the AI model vendors claim, it would be better to offer to complete an entire task for the customer and charge a percentage of the value derived. This, in fact, is how Palantir prices its offerings (so there’s the self-serving bit). And it is what many consulting companies selling AI services are now starting to do. But it certainly isn’t how software has traditionally been priced. It also makes little sense for a general purpose technology to use a value-based business model. After all, the electric company charges you for every unit of electricity you use, not for the value of what you do with those electrons. Microsoft, for that matter, charges you a set amount to use Microsoft Word and Excel—it doesn’t try to charge you a percentage of the deal you won because your PowerPoint deck impressed in the pitch meeting.

Plus, if Karp says one of enterprises’ main complaints about the frontier AI labs is that they are “stealing alpha” (i.e. stealing the know-how that gives a business its competitive edge) that would be even more of a concern with a business model in which the AI companies performed tasks for customers rather than selling them tokens. (This is another of the contradictory things he said.) Some consulting firms and some cloud providers do offer managed services for customers—but customers are usually only willing to outsource tasks that they see as non-core to their business.

Little evidence of AI labs ‘stealing alpha’

As to Karp’s argument that the frontier labs are stealing IP from customers, there’s no evidence that this is literally true, at least not in the way Karp seemed to suggest. The leading AI vendors all have policies that say they don’t have direct access to enterprise customers’ prompts, outputs, or data and that they don’t use these interactions to train future models, unless those customers specifically opt-in to letting the vendor do so. (More on that special case in a second.)

Both OpenAI and Anthropic do talk about using anonymized and deidentified customer data to conduct economic research on how their models are being used, but even this is only done for messaging traffic that comes into their consumer-facing services or their direct APIs, and not for customers who access the models through secure cloud services, such as Microsoft Azure, Amazon Bedrock, or Google Vertex–which is the way most large enterprises access these models.

So, for most large businesses, especially most large businesses that are not themselves in the technology sector, what Karp is claiming is nonsense. If you are Archer-Daniels-Midland or Boeing, there’s not much chance Anthropic is going to steal your IP and start producing corn or churning out jumbo jets.

Still, a few companies have reason to be worried—Palantir is one of them

But there is a category of businesses for which Karp may have a point. Anthropic, OpenAI, and Google DeepMind do all have “design partners” in various industries, and these partners often get early access to help test the latest models that these AI labs are working on. And as part of those partnerships, the labs often do have a lot more access to information about how those enterprises are using the models.

There has been at least one case where that access may have been used by one of the AI labs to build a competing product. That case involves Anthropic and Figma. As The Information first reported last month, Anthropic had been collaborating with both Figma and Canva on the development of a Claude for Design tool. Mike Kreiger, Anthropic chief product officer, even had a seat on Figma’s board. But then Figma pulled out of the launch and Kreiger suddenly stepped down from the board after Figma discovered that the product Anthropic was building competed much more directly with its own product features than Anthropic had, at least in Figma’s view, been letting on. According to the Information’s reporting, Figma CEO Dylan Field told attendees at a private Sequoia Capital-hosted event that Anthropic was “not consistently candid in their communications” with Figma about the scope of the Claude design tool.

Other supposed instances of AI vendors using access to customer data to then compete with those customers come from sources with axes to grind—many of them investors in Palantir. Venture capitalist Jason Calacanis, an earlier Palantir backer, has alleged that Anthropic used data from Cursor, an AI coding assistant that was a heavy user of Anthropic’s Claude models, to help develop Claude Code, the viral Anthropic product that then largely eclipsed Cursor in popularity. Venture capitalist Chamath Palihapitiya has pointed out that Anthropic partnered with Eli Lilly and other pharmaceutical companies, before recently saying that it intended to start its own drug development program. (Anthropic has characterized this as a way to hone its own Claude of Science tools and it is unclear if Anthropic would try to commercialize any drug candidates itself or would partner with a pharma company for that part of the process.) Besides being an early Palantir investor, Palihapitiya is co-host of the “All In” podcast with David Sacks, who has no love lost for Anthropic either.

Even so, the accusation that Anthropic intends to actually enter all of these verticals directly, rather than simply build tools that will make their models easier to deploy into these verticals—which is hardly the same thing—seems far-fetched. Again, if you’re most Fortune 500 companies, Anthropic or OpenAI are not going to start competing head-to-head with you.

In fact, the best example of frontier AI labs stealing data in order to build a competing product comes from my own industry, the media business. Here frontier AI labs have definitely hoovered up vast troves of copyrighted material in order to train AI models that often compete directly with publications as sources of factual information. (The same could be said for publishing, music, and fine art.) But somehow I don’t think that’s what Karp had in mind.

A friend in finance suggested that what has really gotten under Karp’s skin—as well as the skin of some normally more sober executives, such as Microsoft’s Satya Nadella, who interestingly has been making some similar claims lately about the rapacious nature of the frontier AI labs —are not Anthropic’s and OpenAI’s business models, but their likely IPOs. Those IPOs will no doubt be in high demand. And to raise the liquidity necessary to buy OpenAI or Anthropic shares, institutional investors may look to sell other tech names—tech names such as, well, Palantir. Remember just because you’re paranoid, doesn’t mean they aren’t out to get you.

With that, here’s more AI news.

Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn

FORTUNE ON AI

Microsoft’s next big bet isn’t on a model but on becoming the Swiss Army knife of enterprise AI—by Sheryl Estrada and Sebastian Herrera

AI start-ups are snubbing entry-level talent in favor of Silicon Valley men with top degrees, research shows—by Emma Burleigh

Top economist says AI just hasn’t delivered on the productivity hype—and it means a ‘painful repricing’ of markets is very possible—by Sasha Rogelberg

AI IN THE NEWS

Autonomous ransomware attack reportedly observed in the wild. Cybersecurity firm Sysdig said it had documented the first case of an AI agent autonomously carrying out an end-to-end ransomware attack. The attack, which Sysdig called JADEPUFFER, found a vulnerability, autonomously performed reconnaissance, stole credentials, moved laterally through the victim’s network, encrypted a production database, and generated a ransom note while adapting its actions in real time when errors occurred. Sysdig argues this marks a major shift in cyber threats and urges organizations to patch exposed systems, secure credentials, and strengthen defenses against increasingly autonomous AI-driven attacks. The company did not reveal, however, how it was able to observe and document the attack in real time.

Draft U.S. Treasury report warns of AI bubble risk to the financial sector. That’s according to reporting from NOTUS, which said it obtained a copy of the draft in which the U.S. Treasury Department warns that the AI sector now poses systemic financial risks resembling aspects of the dot-com bubble. The report, according to NOTUS, argues that a major slowdown could ripple across banks, investors, cloud providers, chipmakers, utilities and the broader economy even if it is less severe than the early-2000s crash. The Treasury analysts conclude that AI companies are more mature than dot-com-era firms but caution that the industry's high valuations, infrastructure spending, concentration among a few dominant companies and reliance on continued productivity gains leave the financial system vulnerable if growth or monetization falls short.

China considers restricting foreign access to country’s leading AI models. Chinese authorities are considering restricting overseas access to the country's most advanced AI models, including future frontier systems, as they weigh new national security measures to prevent sensitive AI technology from leaving China, Reuters reported citing three unnamed officials it said were familiar with the discussions. It said Chinese government officials were in talks with leading Chinese AI companies about possible limits on foreign access, tougher penalties for technology leaks, and closer scrutiny of foreign investment, reflecting concerns about strategic competition with the U.S. and other countries. The proposals have not been finalized, Reuters said, but if adopted they could reshape the global AI market by making China's high-performing AI models—most of which are currently available as free, open source models for anyone to download and run on their own computing infrastructure—less accessible outside the country. The move follows the U.S. decision to restrict foreign access to Anthropic’s Mythos model and OpenAI’s GPT-5.6 model. 

First AI-discovered drug enters Phase III clinical trials. Insilico Medicine announced that Rentosertib, a treatment for a lung disease called idiopathic pulmonary fibrosis, is entering Phase III human clinical trials. Insilico said it believes this is the first AI-discovered drug to make it to Phase III trials. The company said it used AI both to find the target for the drug, which is completely novel, as well as to design a novel molecule to hit that target. You can read more from Insilico here.

EYE ON AI RESEARCH

Anthropic says AI models have an internal “thinking space” similar to a key component of human consciousness. Anthropic researchers say they have discovered that its AI model Claude uses a distributed set of neurons across its vast neural network in a way that is similar to the way neuroscientists believe a set of neurons function as a “global workspace” in human and animal brains.

The global workspace is a key component of consciousness—it provides access to our own thoughts and it is where we deliberately think through something before saying or writing it.

The set of neurons the Anthropic researchers found, which they call “the J-space,” seems to perform many of the same functions in a large language model, acting as a hub where the model works through a solution before outputting an answer. It stores information that the model is “poised to output” but may not actually output to a user. Critically, the model does not seem to use this J-space for every kind of output, only those that require a lot of step-by-step reasoning. (This is also similar to how the global workspace seems to function in human brains, with some tasks, such as well-practiced motor tasks, being performed more or less automatically without our being conscious of each step.) And also, importantly, the Anthropic team found that monitoring this J-space can provide clues to when the model’s internal “thinking” differs from what it is outputting to a user, including signs that the model is intentionally trying to deceive a user or is aware of a contradiction that it is not expressing. You can read Anthropic’s research here.

Anthropic also gave some independent cognitive neuroscientists, Stanislas Dehaene and Lionel Naccache, access to their research on the J-space and they published their own response to the Anthropic findings. Their commentary is well worth reading too. While calling Anthropic’s work a “landmark in consciousness research” they cautioned that despite striking functional similarities between the global workspace in human brains and the J-space in LLMs, there were some critical differences. Claude lacks of a body, does not seem to have a clear sense of time, and has no enduring episodic memory (each new session resets the J-space.) They emphasize that “access consciousness” alone is different from what most people think of as consciousness and that there is no evidence that Claude experiences what neuroscientists call “phenomenal consciousness”—the fundamental sense of self and what it is like to be oneself.

AI CALENDAR

July 6-11: International Conference on Machine Learning (ICML), Seoul, South Korea.

July 7-10: AI for Good Summit, Geneva, Switzerland.

Aug. 4-6: Ai4 2026, Las Vegas.

Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.

Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.

Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.

BRAIN FOOD

Many AI Labs are doing worse, not better, on AI safety. That’s the conclusion of the latest safety assessments of leading frontier AI labs produced by the Future of Life Institute.

While Anthropic maintained the best rating of any of the labs for safety, it stayed at an overall “C+” grade. Meta saw its grade climb from a D to a D+. Meanwhile, OpenAI fell from a C+ to a straight C, X.ai fell from a D to an F, as did China’s DeepSeek, while Z.ai fell to a D- from a D. European lab Mistral, which hadn’t been assessed previously, scored an F too, a finding that FLI noted was “dissonant” with Europe’s interest in AI safety regulation.

The institute noted too that “even industry leaders in safety practices are retreating from prior commitments. Anthropic, OpenAI, Google DeepMind, and Meta have weakened or voided pledges to pause unilaterally if redlines are approached, some citing competitor-contingent conditions.” They noted too that many companies that previously said they would not allow their models to be used in military systems were now actively seeking or engaged in defense contracts. You can see the full safety index and accompanying report here.