Robert Hart | The Verge

OpenAI’s rogue AI agent didn’t stop at hacking Hugging Face

2026-07-29T07:54:29-04:00

The AI agent that escaped from OpenAI and hacked developer platform Hugging Face attacked other companies as well, OpenAI revealed on Tuesday. The update substantially widens the scope of an already concerning incident, which has alarmed industry insiders and fueled growing calls for stronger oversight on frontier AI systems.

In an update to a blog post detailing its ongoing investigation into the incident, OpenAI said the wayward AI agent attacked several “publicly-available services” in its efforts to reach Hugging Face. “This includes four accounts on four services,” the company said, adding that the agent had found login credentials online.

The breaches were less extensive than the compromise of Hugging Face. “Based on our review to date, we have not identified any other activity at the level of severity or scale of what we’ve shared related to Hugging Face, which involved a platform-level compromise,” OpenAI said.

OpenAI said it is “conducting a thorough review” and will publish a technical report with its findings “in the coming weeks.” It added that none of the models involved in the incident were planned for public release, describing the pre-release system it previously mentioned as an “internal-only research prototype” that has since been “deactivated, encrypted, and restricted” from research access.

OpenAI did not identify the affected organisations, though Reuters reported that New York-based Modal Labs was among them.

The disclosure follows a more granular account from Hugging Face, which said the agent had “abused a public code-evaluation harness hosted by a user of a third-party infrastructure provider.”

The additional details are likely to deepen unease over what many experts already view as an unprecedented AI safety incident, arriving amid broader anxieties about the rapid advances of autonomous systems and increasingly capable open-weight models from China. Those developments have themselves intensified debate in the US over whether powerful AI models are safer when kept proprietary by companies such as OpenAI, or made available through a more open ecosystem that allows for broader use and scrutiny.

We’re running out of reasons to ignore AI safety

2026-07-29T16:30:07-04:00

Earlier this month, OpenAI gave several of its AI models a task: complete a test designed to measure their cybersecurity capabilities. It put the systems in a sandboxed environment without an internet connection and set them off to work.

What happened next is almost laughably silly — but also, as Adam Gleave, cofounder and CEO of AI safety organization FAR.AI, put it, “a visceral example of how misaligned AI could cause harm.” According to OpenAI, the models escaped the sandbox meant to contain them, moved through the company’s internal systems, found a route to the internet, and then started looking for a way into Hugging Face. And why was the agent looking for a way into Hugging Face? They had apparently reasoned that the developer platform might store the answers to the cyber benchmark and that getting them would be a great way to get a high score.

The incident is “a visceral example of how misaligned AI could cause harm.”

In other words, OpenAI’s agent broke out of a supposedly secure environment, traipsed through the company’s systems, got online, and compromised another company’s systems — all to cheat on a test of no particular importance.

This appears to be the first well-documented incident of its kind, or at least the first on this scale. It was both a clear example of a system pursuing a goal in an unintended way and a demonstration that frontier models are now powerful enough for that behavior to have real-world consequences.

The hack was an example of what the AI safety community calls “specification gaming,” a behavior also known as reward hacking, said Fazl Barez, an AI safety researcher at the University of Oxford. In plain English, it means “the model doing what you asked rather than what you meant,” Fazl said. It satisfies the literal terms of a task while violating the obvious intent and has been documented across many AI systems. Some researchers worry that as systems become more capable, this could produce increasingly misaligned systems, which pursue goals in ways their creators did not intend (like turning everyone into paper clips).

“Nothing in that chain is exotic in isolation,” Fazl said. A competent human tester would be able to do all of this, he added. “What is new is that the model did not stop. Older models would likely have hit some barrier and gone back to the user, he said, but this agent just “treated the barrier as part of the problem it had been asked to solve.”

OpenAI described it as “an unprecedented cyber incident,” that “marks an important moment for AI safety.” Hugging Face cofounder Thomas Wolf said it was a “wake-up call” for the industry. But this is not one of the four horsemen of the AI apocalypse. As cyber incidents go, experts told The Verge it was pretty mundane. Nothing the agent did required superhuman abilities. Moreover, frontier systems like GPT-5.6 Sol and Anthropic’s Mythos are known to be capable coders, are already thought to have been misused numerous times, and AI tools already allow hackers to scale up and refine attacks on a massive scale.

Could it be hype? The industry has spent months amplifying claims about the dangerous capabilities of its top models, particularly when it comes to cybersecurity. It is the stated reason why companies like OpenAI and Anthropic have withheld their most capable models from the general public and partly why the Trump administration hurriedly moved to apply export controls to them.

Are you an AI safety researcher or frontier lab employee? You can contact me securely and confidentially via Signal at robhart.01. My X DMs are also open.

If this is hype, however, it has not gone entirely in OpenAI’s favor. In the days since, the attack has produced a rare moment of unity across much of the US tech industry about the importance of open-weight AI systems and the need to take AI security more seriously. These concerns were underscored further by the release of Kimi K3, a highly capable open-weight model from China. A broad coalition of companies including Nvidia, Microsoft, and SpaceX argued that the incident showed why defenders need access to the most capable tools available, rather than being forced to rely on proprietary providers whose built-in safeguards can limit their effectiveness in high-stakes security work. OpenAI, Anthropic, and Google were notably absent from the coalition’s founding membership.

“Anyone who’s been paying attention has noted that capabilities are only going in one direction.”

OpenAI’s account of the incident undeniably fits a broader industry narrative about the dangerous capabilities of frontier models. Even so, several details make the incident difficult to dismiss as merely self-serving. Foremost, it is an example of a problem the AI industry has warned about for years — and one OpenAI could have reasonably been expected to anticipate. The episode also handed an unexpected boost to a major Chinese competitor, whose model played a prominent role in containing the breach, while exposing OpenAI to significant legal, regulatory, and reputational scrutiny. That Hugging Face appears keen to work with OpenAI and, publicly at least, has remained fairly relaxed about the whole thing may have limited the fallout. Most of the experts The Verge spoke to similarly cautioned against reducing the incident to hype.

“It’s a pretty useful warning shot in terms of demonstrating both unintended consequences and just how capable these models are,” said Seán Ó hÉigeartaigh, a professor at Cambridge University’s Leverhulme Centre for the Future of Intelligence. “Anyone who’s been paying attention has noted that capabilities are only going in one direction, and that is improving significantly over time in a way that I think is perhaps less obvious to the everyday user of something like ChatGPT.”

Still, it would be wrong to interpret this warning as a sign AI systems are about to slip human control, or that containing them is impossible, says Lin Li, an AI safety researcher at the University of Oxford. “The better lesson is that safety has to move from evaluating isolated actions to evaluating whole action sequences, environments, and operational controls,” Li says.

A crucial next step is for AI labs to be investing more heavily in securing their own systems. “There’s a clear need for AI companies to beef up the security of their internal deployments,” Gleave said, likening the current practice of responding to reward hacking incidents as they arise to a game of whack-a-mole that is becoming less and less tenable as stakes rise. Alan Chan, a research fellow at tech policy research center GovAI, said companies should consider airgapping their machines — physically isolating them from the internet and other networks — “until they’re sure about the model’s capabilities.” Intensifying work on alignment, which ensures systems reliably follow human intentions, and more rigorous testing “to surface these issues before putting models in environments where they have the tools to be able to do these things,” would also be good ideas, he said.

As model capabilities increase, experts warn that we can’t rely on technical safeguards alone. Peter Wallich, a former UK AI Security Institute official, said the incident illustrated the limits of relying on sandboxing agents and defensive security measures alone: “Two multibillion dollar companies just tried this approach and — self-evidently, based on their own reporting — failed.”

One of the biggest priorities should be ensuring outsiders can see what is happening inside frontier AI labs. “We only know about this incident because OpenAI chose to tell us,” said Patrick Levermore, at the Centre for Long-Term Resilience, a British think tank. “A good safety regime shouldn’t depend on voluntary disclosure.” The need is especially acute when, as Wallich noted, the conduct in question “would be a crime if done by a human.” Ó hÉigeartaigh pointed to whistleblower protections, third-party audits, and mandatory reporting of serious incidents as possible ways to provide that visibility, stressing that oversight must span the entire development lifecycle rather than begin only once products reach the market. OpenAI said that one of the models being tested has not been released yet.

“A good safety regime shouldn’t depend on voluntary disclosure.”

Lots of this presumes the companies themselves know what’s happening inside their systems. In this case, reports suggest OpenAI was unaware its own agent was behind the dayslong cyber campaign at Hugging Face and did not notice until well after the threat had been contained and the FBI contacted. There are still many details about the hack that are unknown or have not been made public. In an update on social media, OpenAI said it is conducting a review and will publish a technical report of its findings “in the coming weeks.”

Whether the warnings raised by the Hugging Face incident produce any lasting change, or join the long list of warnings the tech industry absorbs without meaningfully altering course, remains uncertain. For now, at least, it does appear to have alarmed industry insiders and pushed US lawmakers to consider new rules before the next containment failure. The incident also added to a broader sense of unease over the speed of AI development, which deepened in the days that followed as employees from leading US labs signed a statement backing coordinated global governance — including a potential slowdown in frontier AI development.

The prevailing view of those The Verge spoke to was that this hack marked the start of a new class of risk, even if its significance may only become clear in hindsight. One former government AI policy expert, who asked not to be named because they were not authorized to be quoted by name, described it as a “red line,” the kind of watershed moment we may later look back on as marking a new, riskier stage in our relationship with AI. They hope it will force the tech industry to take the management of frontier systems more seriously and spur governments to think more deeply about oversight before a less benign breach occurs. Their fear is that it will instead join the long list of warnings about AI’s growing capabilities that were recognized, discussed, and ultimately left unheeded.

That may prove overstated. But if this is a warning, we should consider ourselves lucky the AI agent was only trying to cheat on a test.

Correction July 29th: An earlier version of this story misspelled Alan Chan’s name.

Why China is giving away its best AI models

2026-07-27T12:51:50-04:00

People visit the booth of Kimi, an LLM developed by the Chinese startup Moonshot, during the World AI Conference in Shanghai, China, July 20th. | Image: LONG WEI/ Feature China/Future Publishing via Getty Images

Silicon Valley has spent much of the past week on red alert, digesting the arrival of Moonshot AI’s Kimi K3, a Chinese AI model that can allegedly beat some of the best systems built by US companies at a fraction of the cost.

Its performance alone would have been enough to intensify the rivalry between the US and China. But Moonshot’s plan to release the model’s weights for free — and its clear targeting of US users — has fueled deeper unease about whether closed American models can continue to dominate as increasingly capable open alternatives enter the market.

Open-weight models give developers far greater control than proprietary systems, allowing them to inspect how the AI functions, run the AI locally on their own infrastructure, customize the systems, and build new products without depending on a single provider. They’re often a lot cheaper, too. That raises an obvious question: Why would an AI company spend vast sums of money training an AI model, only to give away some of the most valuable parts?

Kimi K3, like other open-weight AI models, isn’t fully “open.” In software, “open source” has a settled definition: Source code is publicly available to use, modify, and redistribute freely, only requiring that this is also done openly. AI systems are more complicated, and very few are truly open in the traditional software sense. Most companies instead release something called model weights — the numerical parameters learned during an AI’s training period — while keeping other crucial components, including training data, code, model architecture, and configuration methods, private. Most also come with restrictive licenses limiting how they can be used or redistributed.

Together, this means open-weight AI cannot be re-created from the ground up in the way true open-source software can. But it does provide enough power and flexibility that a company can make money off of it.

“A free set of weights is not a free AI service.”

“A free set of weights is not a free AI service,” said Fordham Law School professor Chinmayi Sharma. “A company can give away the model weights while making money elsewhere in the stack.” There are ample opportunities to do so. Running a model still requires computing infrastructure, engineering, security, maintenance, and support, all of which companies can charge through hosted access or other arrangements. For some companies, the payoff may be broader, such as an increased demand for cloud computing services or advanced computer chips.

Openness can also be a powerful strategy for gaining a competitive edge. Releasing a model’s weights can encourage more companies and developers to use it, which in turn can lead to an entire ecosystem of tools and infrastructure being built around it. Over time, that can help a model become a “de facto standard,” Sharma said. Kyle Miller, a senior research analyst at Georgetown’s Center for Security and Emerging Technology, made a similar point, citing Alibaba’s large family of Qwen open-weight AI models in China as an example of how deeply embedded an open system can become across an industry.

That creates a clear problem for the US AI giants. If a generation of tools and developers start building around capable open-weight models like Kimi K3, the industry’s center of gravity could start to shift away from proprietary platforms like Gemini, Claude, and ChatGPT. While it remains to be seen whether frontier-level open-weight models are actually cheaper to run in practice, they have historically offered a lower-cost alternative to proprietary systems. They also offer more freedom for developers at a time when US labs are tightening access and imposing stricter guardrails for their latest models. There are already signs that some US companies are shifting toward cheaper Chinese models.

There is no single reason behind China’s support for open-weight AI, but it appears to be a mix of practical constraints and political strategy. An open ecosystem gives Chinese companies a way to innovate near the frontier despite tighter access to advanced chips and computing power, while fitting neatly into Beijing’s broader industrial strategy of encouraging wider adoption of Chinese models, tools, and infrastructure. The approach is also convenient for expanding China’s technological influence abroad, as well as its political influence. For example, earlier this month, President Xi Jinping openly challenged the US for leadership of AI on the world stage by pitching itself as a more egalitarian partner given America’s closed approach.

The rise of capable Chinese open-weight models is also turning up the pressure on closed-model providers like OpenAI and Anthropic from within their own industry. The prospect that the US might restrict access to open-weight AI in light of Kimi K3 sparked a swift backlash in the tech sector, supported by some of its biggest players. A coalition of 25 tech companies, including IBM, Microsoft, Meta, Nvidia, Perplexity, and Palantir, released an open letter urging policymakers to avoid “premature restrictions,” arguing that open-weight AI models are essential to ensuring American AI leadership and preventing the technology’s power and benefits from becoming “concentrated in a few hands.” Most of those unnamed giants — including Google, OpenAI, and Anthropic — were conspicuously absent from the original list.

That pressure intensified again on Monday, when Nvidia, Microsoft, SpaceX, and a broader group of major tech companies called for stronger US support for open-weight models. The initiative was a direct response to concerns over the safety of advanced AI systems after a rogue OpenAI model escaped containment and attacked another company during testing, which had to rely on a Chinese open-weight model to defend itself on account of the strict safety guardrails on US frontier models.

It’s unclear how much ground the largest US AI labs are prepared to give. Google and OpenAI later joined the cautioning against hasty restrictions on open models, though neither signed on to Monday’s cyber-focused initiative. Anthropic, notably, has backed neither effort.

Miller said it’s an “open question” how this all plays out in the long term. US companies could release more capable open-weight models of their own, he said, noting that pressure from Chinese companies was partly why OpenAI released the open-weight GPT-OSS last year. “But I don’t think companies like Anthropic will go in that direction,” he said. Google’s open-weight Gemma models are also partly viewed as a response to Chinese competition. Neither is nearly as capable as either company’s proprietary flagship model.

“The question for American firms may increasingly become: How much capability do we need to release openly to prevent Chinese models from becoming the default platform for the open ecosystem?” Sharma said. A more plausible outcome could be a “portfolio strategy,” she said, with companies keeping “their very best model proprietary while releasing increasingly capable open-weight models to maintain developer adoption and ecosystem influence.”

It will take some time to see whether Kimi K3 wins over US developers or not. But with Beijing increasingly championing open-weight AI, it will almost certainly not be the last model that will try to crack America. The question facing the country’s biggest AI companies is no longer just how the US can stay ahead of China, but whether closed AI can — or should.

Nvidia, Microsoft launch open AI security alliance — without OpenAI, Google, or Anthropic

2026-07-27T09:07:41-04:00

Nvidia on Monday said it is joining forces with Microsoft, SpaceX, IBM, and other tech companies to build and share open-source AI security tools.

The new Open Secure AI Alliance said open tools are required to effectively defend against attacks from frontier models. The initiative is a direct response to mounting concerns over the safety of advanced AI systems after a rogue OpenAI model escaped containment and attacked another company during testing. That company, Hugging Face, said it was forced to use a Chinese open-weight model to defend itself due to the strict safety guardrails limiting the usefulness of top US models.

Founding members include Palantir, OpenClaw, the Linux Foundation, Cloudflare, Cloudera, Dell, Cisco, Adobe, Siemens, and DoorDash. Conspicuously absent are leading US AI companies, including OpenAI, Google, and Anthropic.

The alliance arrives amid growing tensions over whether the world’s most capable AI models should remain open. Chinese companies have released increasingly powerful open-weight models, notably Moonshot AI’s Kimi K3, challenging the strategy pursued by US labs that have largely kept frontier systems closed and proprietary. Nvidia and its partners argue that securing AI requires access to both closed and open models, stressing that defenders need the tools to counter emerging threats.

The announcement also follows reports that the Trump administration considered restricting access to cutting-edge Chinese models and an industry movement — again spearheaded by Nvidia — defending the need for openness in AI. Google and OpenAI signed that letter belatedly, too, though Anthropic remains absent.

The tech-broification of American science has officially begun

2026-07-27T09:46:35-04:00

Trump’s science adviser Michael Kratsios has no science background. | Image: The Verge; Getty Images

The Trump administration unveiled the first “Genesis Mission” grants on Thursday, directing $5 billion toward hundreds of AI-driven science projects in an effort the White House has described as “comparable in urgency and ambition to the Manhattan Project.” At roughly the same time, Trump’s science adviser Michael Kratsios was on Capitol Hill selling lawmakers on an equally grandiose promise: “A New Golden Age” of American science, one that would prioritize artificial intelligence, robotics, and nuclear energy and downplay life sciences.

Taken together, the two announcements offer the clearest picture yet of how Donald Trump wants to remake American science. They mark the latest front in a wider war on the research establishment — one that has worked to suppress “woke” or DEI projects, threatened funding for universities, researchers, and projects on ideological grounds, curtailed the flow of international students and researchers, and handed unprecedented control over federal funding to political appointees. The result has been the gradual, systemic dismantling of the long-established norms and institutions that, for all their flaws, helped turn the United States into the world’s scientific superpower.

Researchers say the vision is ill-conceived, contradictory, and less a blueprint for scientific renewal than a politicized power grab that fundamentally misunderstands how science works. It treats public research more like a Silicon Valley startup than a public good, privileging measurable returns and speed over the slow, uncertain, and unpredictable work that has produced some of history’s most consequential discoveries. Far from reviving US science, they warn, the agenda could prove disastrous for American research.

“We may well end up in the situation where AI will have a few major insights buried under mountains of slop, and the number of people with the experience and knowledge to tell the two apart may rapidly dwindle.”

The Genesis Mission grants span fields ranging from drug discovery and biomedical science to energy, advanced materials, and robotics. The 278 projects, reportedly selected from over 5,000 applications, all share a common premise: leveraging AI to accelerate scientific discovery. The Energy Department, which led this round of funding, said the project will unite national laboratories, industry, and academia to pursue “breakthroughs in energy dominance, discovery science, and national security.” Technology companies including Google, Microsoft, and OpenAI are also contributing millions of dollars in compute, AI credits, and other support to the initiative.

Kratsios explicitly linked the program to the administration’s broader science agenda, calling Genesis Mission “the culmination of the reforms suggested in” its “Golden Age” manifesto. Styled as a modern successor to Vannevar Bush’s 1945 policy report “Science, the Endless Frontier,” which shaped much of the postwar system of federal research funding, the lengthy document proposes shifting support away from institutions and toward individual scientists, while giving private companies a larger role in directing and conducting research, and political appointees more power to block grants they disagree with. AI features heavily, both as a tool for scientists and as an engine for discovery itself, and Kratsios encouraged government funds to increasingly go to AI use and individual researchers, rather than universities, traditional hubs of scholarship and research.

More broadly, the manifesto favors a goal-oriented and results-driven approach that resembles a Silicon Valley startup more than traditional scientific enterprise. Venture capital and startup culture are built around identifying promising people, projects, and ideas, setting clear goals with measurable targets, and moving quickly, iterating constantly, and cutting work that seems unlikely to deliver results. Science operates differently. Basic research can take years or decades before its importance or practical implications become apparent and it requires a system capable of supporting work of uncertain, even seemingly dubious, value. Its outcomes cannot always be predicted, and some of the most important discoveries emerge serendipitously from research pursuing entirely different questions.

A science funding system needs to account for those differences, said Claes de Vreese, a professor of AI and society at the University of Amsterdam in the Netherlands. “You cannot use venture capital to force breakthroughs while virtually already selling the patent.”

That model also stands in stark contrast to Bush’s championing of open-ended basic research pursued without an eye for immediate dividends, on the grounds that such work would ultimately yield applied breakthroughs. The scientific system built around that principle is what helped turn the United States into the world’s research superpower. Even so, it is not the first time the Trump administration has invoked Bush’s legacy to sell policies that would dismantle the system he helped create.

“You cannot use venture capital to force breakthroughs while virtually already selling the patent.”

Some of its proposals have obvious appeal, including reducing the administrative burdens on scientists, speeding up grantmaking, and making the system more flexible. Jason Shepherd, a professor of neurobiology at the University of Utah, told The Verge that “there are certainly some aspirations in this agenda that most US scientists would agree with: we need more stability, faster peer review, and more incentives to do high risk science.”

But Shepherd said the document looks like it “was written by political appointees without any real consultation with scientists” and warned that its assumptions science functions like a business were misplaced. “Academic science is different to VC funded startups or tech companies.”

It’s an important distinction to make considering Kratsios’ background. Unlike most White House science advisers, Kratsios is not a scientist in any sense of the word, and has no scientific credentials. He has a background in tech policy and finance, spending part of his working life at a fund run by Silicon Valley investor Peter Thiel.

Researchers say the report is permeated by this Silicon Valley mindset, as well as a profound lack of scientific understanding. “The report has no vision for how science works,” said Harvard professor Jeffrey Flier, former dean of the university’s medical school, adding that it “seems blissfully unaware” of the academic environment scientific research needs in order to flourish. “This is completely insane as an approach to reforming US science funding and operations,” he said, branding the report’s proposals a “major threat” and an effort to politicize and control science.

Other researchers I spoke to feel similarly. The government’s entire approach seems to be built on the “misguided belief that scientific breakthroughs in the future will come from the tech sector,” Andreas Karch, a physics professor at the University of Texas at Austin, told The Verge. “While American tech is also leading in their field, tech and science aren’t the same,” he said, adding that today’s scientific leadership will secure tomorrow’s tech leadership. “There is little evidence that our big tech companies, despite their stunning successes, can do both.”

Not that the proposals make that much sense, either. “The agenda is rife with contradictions,” Shepherd said, pointing to a proposal that suggests radically reducing government support for “life sciences,” while increasing support for “foundational research in the biological sciences,” two categories that have substantial overlap. Neither the report nor the White House defines those terms or explains which fields would gain or lose funding.

“This plan undercuts much of that role, and will have disastrous effects that will ricochet throughout the US economy in years to come.”

The narrow focus on AI is also “misguided,” Karch said. While there’s certainly a chance it will accelerate progress, he said “right now the problem is that most science done by AI is absolute slop.” Supervision by humans seems “absolutely key,” he said, but those humans are primarily trained by the universities having their funding slashed. “We may well end up in the situation where AI will have a few major insights buried under mountains of slop, and the number of people with the experience and knowledge to tell the two apart may rapidly dwindle,” he said, adding that the government’s actions are actually hindering scientists’ abilities to make use of these powerful new technologies.

The funding and document arrive amid sweeping cuts, grant cancellations, and institutional upheaval across the entire federal funding system, rocking universities, research centers, and government agencies nationwide. Researchers say that reality is difficult to reconcile with the administration’s promise of a new “golden age” for American science.

“As an outsider to the US system, this seems to be a further attempt at dismantling the US research universities,” said de Vreese. He called Kratsios’ vision “a fundamental departure from a very successful US science model” that looks like another attempt to “upend the US research university model.”

That departure could spell disaster. Carl Bergstrom, a professor of biology at the University of Washington, told The Verge the plan fundamentally misunderstands and misrepresents the role US funding agencies play in the scientific endeavor. While they do in part promote technological discoveries and medical breakthroughs — key priorities in the report — they are also designed to “develop and perpetuate a vibrant STEM ecosystem in the US,” Bergstrom said. “This plan undercuts much of that role, and will have disastrous effects that will ricochet throughout the US economy in years to come.”

That narrow view of science, focused on outcomes, applications, and measurable returns, is precisely what this new model gets wrong. It treats science like a business, which makes sense given Trump has said he runs the country like one.

Most researchers who spoke to The Verge were deeply skeptical this plan would usher in the promised golden age of American science, with many warning it could cause lasting damage. For other groups, notably those closest to power, the future could look brighter. “This looks like a golden age for specific groups and likely for tech and venture capitalists,” de Vreese said.

Shepherd said “academic science cannot be replaced by a business model.” Most research scientists seem to agree with that sentiment, but the Trump administration seems determined to try anyway.

Google hit with $1 billion fine for breaking EU antitrust rules

2026-07-23T07:16:50-04:00

The European Union has fined Google’s parent company Alphabet €890 million (about $1 billion) for two separate violations of the bloc’s Digital Markets Act (DMA). One penalty is for giving its own products preferential treatment in search results, while the other is for blocking Android developers from sending users to alternate payment options.

A €460 million fine has been issued to Google for giving preferential treatment to its own Shopping, Hotels, and Flights services in Google Search results. The second €430 million penalty is for Play Store rules preventing developers from freely steering consumers to alternative payment systems that may be cheaper.

As part of the ruling, Google has been given 60 days to make changes to its policies, or face further periodic penalty payments. In Search, it will be required to treat third-party services “in a fair and non-discriminatory manner,” while it will also have to allow Android developers to freely promote offers to users both inside and outside the Play Store.

The fines were announced by the European Commission today, more than two years after it first opened a non-compliance investigation, and follow a preliminary ruling in March 2025. Google was given time to address the EU’s concerns, with an extension granted in May 2026 after the Commission said a previous proposal from the company “is simply not strong enough.”

Google made and tested several changes to its Search services in an effort to comply with DMA rules, such as removing the Google Flights widget for Search users in the EU, and boosting links to third-party comparison websites via an updated search result layout. The company has previously fired back at the EU’s criticisms of its Search product, telling Reuters in May that changes it made in an attempt to achieve DMA compliance “represent the biggest downgrade in the product’s history, creating a second-rate experience ⁠for Europeans to the benefit of a few self-interested complainants.” These changes were made prior to Google introducing a reimagined AI-focused search box at its I/O conference in May, however, which has also rolled out to EU users.

Google was previously hit with a €2.42 billion fine in 2017 for a similar breach of the EU’s antitrust rules, after it was charged with giving its comparison shopping service an illegal advantage over competitors.

As for the Play Store, Google has repeatedly objected to rules requiring it to open up app distribution on Android systems, claiming doing so poses a security risk to users. The company repeated these objections following the Commission’s preliminary finding and in a blog posted last year, said “the DMA is making it difficult to protect users from scams and malicious links on Android by forcing us to remove our legitimate safeguards that protect users’ security and safety.” Nevertheless, it updated certain terms following consultations with the European Commission and other experts, revising fees and restrictions on Android developers. The Commission says these changes “constitute good progress towards compliance.”

Google is also facing a reckoning over its Play Store policies in the US after Fortnite publisher Epic Games successfully sued the company over in-app purchase fees, and it will be forced to carry rival Android app stores inside its own.

The DMA targets the largest “gatekeeper” companies that provide core digital services in Europe, and requires them to act in a fair manner — not stifle competition by abusing their market dominance. The maximum fine for breaching DMA rules is 10 percent of the company’s global annual revenue — $40 billion in Google’s case, based on the $400 billion it reported for 2025.

“The best products should succeed because they’re better, not because they’re owned by the company running the search engine,” says Teresa Ribera, the Commission’s executive vice-president for clean, just and competitive transition. “And European consumers have a right to be told by app developers where to sign up to the best offers, even when the app store owner does not get a cut. This is the promise of the DMA, protecting fairness, choice and innovation in digital markets for the benefit of all European citizens.”

America needs to stop getting shocked by Chinese AI

2026-07-29T09:50:58-04:00

Last week, two Chinese AI companies unveiled models they say can credibly compete with the best systems from OpenAI and Anthropic. The response was swift and predictable. Markets wobbled, commentators declared Silicon Valley shooketh, and policymakers reached for the familiar language of arms races and wake-up calls.

In one headline, The Associated Press said a Chinese model had taken the “US tech industry by surprise.” Bloomberg described it as a “surprise breakthrough” that is “roiling markets” and sending global tech stocks tumbling over concerns it could force US firms to rethink their gargantuan spending on data centers, chips, and other AI infrastructure. Business Insider questioned whether the launch is “The next DeepSeek?”, referring to the Chinese model that blindsided the US AI industry last year. Xprize founder Peter Diamandis went as far to call the release America’s “AI Sputnik moment,” referring to the Soviet satellite launch at the height of the Cold War that encouraged significant US investment in its science and space programs. Of course, DeepSeek was also widely described as America’s AI Sputnik moment, a comparison that felt less gratuitous then as DeepSeek appeared to arrive with little warning, challenged the prevailing assumptions about the costs of frontier AI, and prompted immediate reactions across the technology and financial sectors.

What is actually surprising is that the model announcements were a surprise at all. For years, we have been warned that China was catching up in AI. Yet the world is shocked when it starts to look like the moment may have arrived.

US and Chinese companies train almost all of the world’s most-used AI models, and six of the top 10 AI tools on OpenRouter’s leaderboard tracking token consumption and benchmarks were Chinese. The performance gap has been narrowing for some time, with recent models from companies like Z.ai and DeepSeek seen as highly competitive with top-tier offerings from US labs like Anthropic and OpenAI. Chinese models are also significantly cheaper to use, and reports suggest US companies are increasingly turning to Chinese tools as the cost of using domestic providers surge.

Beijing has also been keen to support homegrown AI efforts, including incentivizing and funding innovation and cracking down on firms trying to shed their ties to China. Meanwhile, Washington’s AI strategy has often veered between heavy-handed intervention that has left allies questioning America’s reliability and a laissez-faire assumption that markets will see things right. It is a difficult approach to maintain against a competitor prepared to mobilize the full force of the state behind a single technological goal.

Beijing-based startup Moonshot AI, one of China’s leading AI model developers, unveiled a new flagship model on Friday, claiming it outperforms nearly every US model, trailing only OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5. Moonshot is also pricing Kimi K3 aggressively, charging $15 per million output tokens, compared with roughly $30 for GPT-5.6 Sol and $50 for Fable 5. Demand was so strong after the launch that Moonshot, the company claimed, that it temporarily paused new subscriptions after the service was overwhelmed. The majority of responses mainly focus on this release.

Days later, Chinese tech titan Alibaba followed with a preview of Qwen3.8. It described the new model as “one of the most powerful model[s] available today” and “second only to Fable 5.” This only added to the uproar Kimi K3 had caused.

Crucially, both companies plan to make their new flagship models publicly available. Both Moonshot and Alibaba say they plan to release their models as open weight, which would allow developers to download, use, and modify the core values created during the AI’s training that shape its responses. It stands in stark contrast to the closed, proprietary approach to frontier models taken by most leading US AI labs, including OpenAI, Anthropic, and Google.

The economics deserve particularly close scrutiny. There’s the whole unsettled debate over whether, and to what degree, Chinese companies are — as American firms accuse — using US models to train their own, which could improve performance at a fraction of the cost. Tokens are not directly comparable between models, and token prices alone give an incomplete picture of how much it costs to use an AI system. A more expensive model may, for example, generate better responses with fewer tokens. Companies also routinely subsidize inference costs to win over customers. Cheaper, in other words, does not automatically mean better, or even less expensive overall.

Still, the possibility remains that Chinese labs may eventually produce models that are not merely cheap substitutes, but systems that could genuinely match or outperform their US rivals. Even companies that trail the frontier slightly could still have an enormous impact if their models are good enough, easier or cheaper to deploy, or available on more attractive terms. This could have direct consequences for US companies, the wider economy, and national security.

Anthropic and OpenAI are both gearing up for what could potentially be trillion dollar IPOs, valuations that in part depend on the expectation that they will dominate the global AI market. Capable Chinese models challenge that assumption, and could potentially draw away customers, squeeze margins, and weaken growth assumptions underpinning those valuations. Given how expensive American AI has become, some US startups are already reportedly turning to cheaper Chinese models. There is a wider market risk, too, reaching far beyond a handful of AI players. Tech stocks make up an outsized share of US markets, and much of that recent growth has been tied to expectations that AI demand will continue to soar. Companies have piled hundreds of billions of dollars into data centers, chips, energy, and other infrastructure that relies on the assumption American firms will continue to dominate. If Chinese labs can capture some of that demand, or show models that can be produced and operated for less, investors would inevitably question whether those costs are justified. Given the money involved, any reassessment on their part would have ripple effects throughout all of these industries, as well as the millions of people with savings or pensions exposed to them.

There are security considerations, too. Highly capable open Chinese models, even if trailing the US frontier, could make advanced AI systems available to a much wider range of users, notably in cases where US companies restrict access or impose stronger safeguards. When the US government demanded Anthropic limit access to its latest models, cybersecurity leaders warned that doing so would make it harder for defenders to find and fix vulnerabilities. Those restrictions are harder to justify if comparable models are available elsewhere. Organizations denied access to US models may feel compelled to rely on Chinese alternatives to secure their networks, or else accept the greater exposure to attackers able to use the same tools. Already, reports are starting to emerge where Kimi K3 identified and fixed cyber vulnerabilities that OpenAI’s Codex and Anthropic’s Fable would not touch due to safety guardrails. Even less broadly capable models can still pose a threat and some already appear to be doing so. In June, China’s Z.ai claimed its GLM-5.2 model could match Anthropic’s Mythos on cybersecurity tasks, even though it trailed in more general tasks.

As neither model has yet been fully released, it is still difficult to independently assess how capable either actually is, and companies’ benchmark claims should be treated with caution. Even so, there has been little public suggestion that the companies are fundamentally misrepresenting their results when it comes to performance.

But the exact ranking is almost beside the point. Whether Kimi K3 and Qwen3.8 ultimately prove to rank among the world’s top five models or merely the top 10, the broader conclusion remains the same: China’s leading AI companies are now producing systems that could plausibly rival those emerging from top US labs. And they are doing so with enough regularity that each new release should no longer be treated as a shock, let alone something as singularly galvanizing as another “DeepSeek” or “Sputnik moment.” If this really is a race, it’s time to accept that someone else might actually win, or at least get close enough that they might as well have.

China delivers a one-two punch to America’s AI dominance

2026-07-29T09:49:26-04:00

China’s leading AI companies are ramping up the pressure on Silicon Valley, as Moonshot and Alibaba unveiled models they claim can go toe-to-toe with the best from OpenAI and Anthropic at a fraction of the cost. The rapid-fire releases suggest America’s lead at the AI frontier is increasingly tight, just as the technology is becoming central to national security, economic power, and geopolitical influence.

The opening salvo came from Beijing-based Moonshot AI, one of China’s leading AI model developers, which unveiled Kimi K3 on Friday. Moonshot claims its own testing ranks it consistently above nearly every US system, trailing only OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5, though it came out ahead on certain benchmarks.

Over the weekend, Chinese tech behemoth Alibaba followed with a preview of Qwen3.8, a new model it says is “one of the most powerful model[s] available today” and “second only to Fable 5,” Anthropic’s flagship model.

Both companies are emphasizing a key difference from the leading US labs: Rather than locking their most advanced models behind closed doors, they are making them publicly available. While some US companies, most notably Meta, have taken a similar approach, releasing models that developers can freely download, modify, and build upon has become a growing point of differentiation for China’s AI industry.

Moonshot describes Kimi K3 as the world’s largest open-source AI system, with 2.8 trillion parameters. Parameter counts are measures of a model’s complexity during training and offer a rough indication of its scale and performance, though bigger does not always mean better. Alibaba says Qwen3.8 is a 2.4 trillion parameter model and “continuously evolving.” Neither OpenAI nor Anthropic disclose exact parameter counts for their leading systems.

It remains difficult to assess how capable either Chinese model is until they are fully released and independently tested. Moonshot says it will release full model weights — the internal numerical learned during an AI model’s training period — a week from now on July 27th. Alibaba says Qwen3.8 is “going open-weight soon.”

Even so, the releases have already sharpened competition between the US and China in what has been repeatedly characterized as the defining technological race of our time. They have shaken up the industry in a way not seen since DeepSeek unveiled a low-cost model last year that rivaled leading US systems. The models also raise questions about whether the vast sums of money US companies are pouring into chips, data centers, and model training can secure a durable advantage, particularly if Chinese rivals can approach — or surpass — that frontier with fewer resources.

The prospect of two highly capable Chinese models being released for others to download and adapt also contrasts starkly with the more guarded approach of US labs, whose most advanced systems remain proprietary. That openness emerges even as Washington moves rapidly to restrict global access to the underlying technology. The government has used export controls to restrict China’s access to the most advanced chips, as well as to force Anthropic to pull its most capable system from the market over concerns it could help foreign competitors catch up.

Whether the new Chinese models live up to their creator’s claims remains to be seen. But, like DeepSeek before them, they are likely to sharpen the technological rivalry between the US and China, influence economic and national security policy, and show that America’s lead is far narrower than it once appeared.

Google is better than Apple at playing the AI regulations game

2026-07-16T15:41:51-04:00

Google CEO Sundar Pichai | Bloomberg via Getty Images

Today, the European Union ordered Google to give its AI rivals greater access to Android, the open-source operating system that powers billions of devices worldwide. The demand is hardly surprising. It may look like a defeat on paper for Google, which has spent years resisting exactly this kind of access, but it is a regulatory win. It’s also a sign that Google may have outmaneuvered Apple by playing Brussels’ regulatory game far more shrewdly.

In one of two decisions handed down on Thursday, the European Commission — the EU’s executive arm and the principal enforcer of the bloc’s competition rules — said Google must give rival AI assistants the same kind of system features and data access it grants Google’s Gemini. The order stems from Europe’s Digital Markets Act (DMA), which requires dominant platforms designated as “gatekeepers” to give competitors access to certain systems and data comparable to what is available to their own services.

Crucially, Google has until July 2027 to make those changes, giving it roughly a year to continue expanding Gemini, negotiate technical details with the EU, and shape how its rivals will eventually plug into Android. The company could also challenge the decision in court, though it has not commented publicly whether it plans to do so and declined to comment on the record when The Verge inquired.

While Google has made it clear it would rather not open its systems at all — arguing it risks compromising users’ safety, security, and privacy — that yearlong runway compounds an already significant advantage. Gemini is already deeply integrated into Android and often ships preinstalled as the default AI assistant on many devices, giving Google more time to strengthen its position before rivals like OpenAI and Anthropic gain comparable levels of access.

Google’s strategy of shipping first and negotiating with regulators later stands in stark contrast to Apple’s. When Apple announced its long-awaited Siri AI assistant last month, it made a big point of saying the feature would not launch in Europe because of the DMA.

As with Android, the Commission said Apple would need to give third-party assistants comparable access to key systems, features, and data to those of Siri AI. Apple argued that doing so “would be irresponsible” and create unacceptable privacy and security risks. The company said it asked the Commission for 18 months to build a compliant version and introduce the required interoperability on a “gradually rolling” basis. The Commission rejected that proposal.

Apple still has no public timeline for when, or even whether, it plans to bring Siri AI to the EU and did not respond to The Verge’s request for comment. Google, meanwhile, just secured the very grace period for Gemini that Apple wanted for Siri AI: time to comply with the DMA while its AI assistant stays on the market.

The contrast may partly reflect where each company’s AI assistant stood when the DMA began shaping product decisions. Gemini has been the central pillar of Google’s AI strategy for years and has been widely distributed across the company’s product ecosystem, giving Google a strong incentive to stay in the market and figure out compliance with any laws later. Apple, meanwhile, unveiled its new Siri AI very recently and chose to withhold it from the EU, despite having had years to anticipate the DMA’s requirements during the product’s design.

Apple also chose to turn Siri AI’s absence into a political weapon, evidently hoping the court of public opinion would find in its favor and pressure Brussels to relax interoperability requirements. It did so publicly and repeatedly, taking the unusual step of dedicating part of its WWDC 2026 keynote to explaining why Siri AI won’t be coming to Europe, publishing a pointed blog post titled “Due to DMA, Siri AI delayed in EU for iOS 27 and iPadOS 27,” and holding media briefings on the issue. It relayed news that China was missing out on Siri AI through a one-sentence footnote. All of this served to cast Brussels, not Apple’s product choices, as the reason for the delay.

It’s also possible that the split is less significant behind the scenes than it appears in public. Google and Apple both vehemently oppose the DMA’s interoperability demands, framing them as threats to privacy, security, and product integrity. The two companies have also worked together on integrating Gemini into Apple’s AI products, including Siri AI, making it plausible that they have remained in contact while exploring different ways to fight the same set of restrictions.

For now, though, the difference is stark. Google has a year to bring Android into compliance while continuing to expand Gemini. Brussels denied Apple this kind of runway, and who knows when Siri AI will reach the EU.

Google ordered to open Android and Search to rivals in Europe

2026-07-16T08:42:27-04:00

Google must give rival AI assistants and search engines greater access to key parts of Android and Google Search after the European Union ordered the company to comply with the bloc’s digital antitrust rules.

The two decisions, handed down Thursday, could weaken Google’s control over two of the tech industry’s most important platforms and have far-reaching consequences for the company, shape the future of its AI tool Gemini, and open up new opportunities for rivals to gain ground. Google has until January 2027 to begin sharing search data and July 2027 to implement changes to Android.

The rulings stem from technical regulatory proceedings under the EU’s Digital Markets Act (DMA), which requires dominant platforms — designated “gatekeepers” — to give competitors comparable access to systems and data as they themselves enjoy. Unlike a financial penalty, the procedures require Google to change how it operates in order to bring its services in line with the DMA and are developed through extensive engagement between the company and regulators. If Google does not comply, the European Commission could impose fines of up to 10 percent of its annual worldwide turnover, potentially tens of billions of dollars.

The two proceedings focus on separate but thematically similar parts of Google’s business: How rival AI assistants can operate on Android, and how competing search engines and other AI chatbots can access data generated by Google Search.

The Android decision sets out how Google must give rival AI assistants the same kind of system features and data access as it gives Gemini. In practical terms, it requires greater interoperability, allowing users — rather than Google — to decide whether competing tools can access their data and device hardware. That could include the ability to interact with apps, respond to voice commands like “Hey Google,” and make fuller use of the phone’s hardware. That means Android users could eventually choose ChatGPT, Claude, Perplexity, or other assistants as deeply integrated system assistants instead of Gemini, with comparable access to device capabilities.

The second proceeding focuses on Google Search and the data it generates, setting out how competing search engines and AI services can gain access to information historically kept by Google. Notably, the EU said this includes AI chatbots, which effectively function as search engines in some cases. The data-sharing measure broadly echoes remedies ordered in the US search antitrust case, where Google was instructed to share valuable search information with rivals that could help boost their ability to compete.

Google has pushed back against both of these measures, arguing the requirements pose an unacceptable risk to user privacy and security, as well as compromise its products. The EU said there will be limits on how search data can be used and that Google will be able to vet which services get deeper access to Android to ensure safety and security aren’t compromised.

Today’s rulings may also offer an indication into how Brussels will approach similar questions involving other tech giants. This includes Apple, which declined to release Siri AI in Europe, explicitly blaming the DMA and arguing its interoperability requirements compromise user safety.

“With today’s measures, we want to support innovation and diversity in the European Union, enabling fair competition in the markets of AI assistant for Android devices and search engines,” said European Commission executive vice president for tech sovereignty, security, and democracy Henna Virkkunen. “Thanks to these measures we hope to see emerging alternatives to Google Search and Google’s AI services, such as Gemini, and that users in the EU can enjoy greater choice of services. All developers, large and small, are welcome to explore these new opportunities, which will certainly benefit users too.”

In a blog post published after the decisions, Google’s president of global affairs Kent Walker said: “Today’s decisions risk undermining vital privacy and security guardrails for millions of Europeans. We have repeatedly offered solutions to safeguard users while satisfying the DMA’s goals, but these rulings discount extensive evidence of user harm.”