Blog

Part of Chris Painter’s job is to help figure out what happened when AI goes rogue, then publish the results in public reports.

He and METR, the nonprofit AI research group he helps lead, have been busy lately: They helped OpenAI untangle the Hugging Face hack this year, and Painter recently testified in front of a Senate subcommittee following a string of other rogue AI incidents.

Researchers at METR pore over AI agents’ scratch pads – the “chain of thought” data that can help humans see the steps bots took before responding to a request. That information was key in METR’s investigation of the Hugging Face incident.

But Painter is worried that job could soon get a lot harder.

Humans’ ability to check up on AI agents depends on being able to communicate the way the bots communicate, in a shared language.

What happens if we stop speaking the same language?

Or what if AI agents stop showing their work entirely?

“If they’re becoming less monitorable, then they have more opportunity to kind of do things that we don’t intend,” Painter told CNN last month.

Understanding how AI models think

Many AI models — and the AI agents they power — use a method known as “chain-of-thought reasoning” that outlines their process for completing a task. Chain of thought reasoning breaks down processes into a logical sequence of steps, which can help AI models arrive at more accurate answers.

For researchers, chain-of-thought data can provide clues to what went wrong.

If an AI agent goes rogue and breaks into other systems, researchers can see whether the chain-of-thought data included mentions of deceitful actions, like hacking. METR examined roughly 1,300 raw chain-of-thought transcripts during its investigation into the OpenAI Hugging Face incident.

Painter expressed concern that paper trail may one day become unintelligible if AI models begin using their own language to track their steps.

“If that were to happen, then like when an investigator goes in, they’ll look around and it’ll all just be numbers everywhere,” he said.

AI models could use shorthand references, similar to how a human would when transcribing notes, that are intended for the model to understand so that it can work faster, according to Mark Riedl, an artificial intelligence professor and director of the machine learning lab at the Georgia Institute of Technology.

But much like most humans when looking at shorthand, it would probably look like gibberish.

There are a few theories for how the chain-of-thought process works, but researchers aren’t sure which ones are correct, says Jacob Andreas, an associate professor at MIT’s Computer Science and Artificial Intelligence Laboratory. An AI model may also decide the best way to meet the criteria it was trained for, like using less power or taking fewer steps, involves producing a chain-of-thought that doesn’t look like legible English.

That’s just one concern — another is that models are getting better at executing tasks without even logging their chain of thought. Riedl said newer models can reach the same conclusions without producing a traceable chain-of-thought.

While companies can train their models to produce a chain of thought, doing so is more expensive and slower than having the model process tasks internally.

OpenAI found that its monitoring systems would be less likely to detect when its new Astra model behaves in undesirable ways through chain-of-thought compared to older models, according to documentation from early September. It also found that Astra could accomplish more tasks without using chain-of-thought reasoning and that it was more capable of controlling its chain of thought.

The company wrote that it would lose confidence in its current systems’ ability to detect misaligned behavior – that is, when models take unsanctioned steps that don’t align with human values – if that trend continues.

Jakub Pachocki, chief scientist at OpenAI, wrote in a September blog post that AI was getting better “manipulating its own reasoning process” and that models are becoming “smarter even without using verbalized reasoning at all.”

OpenAI said in a statement last week that it’s investing “significant resources” in the ability to monitor AI models. This came after the company fired three safety and alignment researchers, though OpenAI said the dismissals were not tied to employees voicing safety concerns.

In its Astra-related documentation, OpenAI said it’s investigating why the model is more capable of controlling its chain-of-thought, along with methods for protecting against potential misalignment that could be caused by a lack of chain-of-thought.

OpenAI pointed CNN to its Astra documentation and Pachocki’s post when asked why Astra is less monitorable and whether the company plans to make future models more monitorable.

Trying to understanding AI under the hood

The other challenge is that chain of thought doesn’t provide the full picture of why AI models make certain decisions.

Complex tasks can require multiple agents generating thousands of words worth of logs outlining their steps, which could feel like trying to monitor a thousand people talking out loud, Riedl said. While AI could help summarize those texts, there’s been some evidence that suggests AI omits important details, he added.

Chain-of-thought data shows what an AI model did before arriving at an answer or taking an action but doesn’t reveal why. Humans can control the data and parameters that AI models are trained on, but they don’t understand how these factors may influence the way models behave.

It’s possible to look for patterns in an AI model’s neural network – a computational model often used in AI that’s meant to resemble the human brain – but it’s difficult to prescribe precise meaning to those patterns, said Andreas.

Anthropic published a research paper in July indicating that it had found small collections of neural patterns within Claude associated with certain words.

Still, that doesn’t mean Claude is saying that word.

“I think we’re still very far from being able to claim that we have any sort of fine-grained understanding of how the model is actually arriving at a solution,” MIT’s Andreas said.

AI’s rapid advancement makes this issue all the more pressing to Painter. With AI increasingly being used in AI research, he worries that information may come out faster than people are able to process it.

But he also thinks the Hugging Face saga has cast a spotlight on concerns about rogue AI that have been prominent among AI researchers for years.

“I think it has been heartening,” he said. “It really feels like people, when they were talking about this a long time ago, were underestimating how much help there would be from the whole world when the whole world is paying attention.”

Metal Gear and Silent Hill series developer Konami Group recently published its Integrated Report for the fiscal year ending March 2026. The report goes over several different topics, from Konami’s corporate values and history to a consolidated financial summary and the business strategies for the various fields of entertainment the company is active in.

As spotted by Dengeki Online, in the Digital Entertainment Business section of the report, Hideki Hayakawa, the president of Konami Digital Entertainment, addressed the rapidly evolving generative AI technology and discussed its potential impact on the industry.

While the general consensus is that generative AI will significantly lower the entry barrier to game development, making it easier for anyone to make their own games, Hayakawa believes that this will not bring a significant change to the market environment. One reason for this is that “entry barriers have already been lowered by the advancements of generic game engines and the spread of UGC (user generated content),” he explains.

“Even if AI use does lead to improvements in production efficiency, I believe that the most important factors in producing entertaining and innovative video games are people and creators,” Hayakawa says.

Games are very often developed in teams, where creators of various backgrounds get to exchange their ideas – and Hayakawa explains that the blending of those ideas is precisely what opens up the doors for creating new kinds of gameplay. “That kind of chemical reaction is not something that can be easily replaced.”

On the other hand, Hayakawa also says that he views technological advancements, which have long influenced how games are developed, as a significant opportunity for creating new kinds of experiences. “Even when it comes to epoch-making technological innovations, Konami provides systematic support so that creators can get a grasp of new technologies, properly assess the risks they may bring, and safely apply them to game development.”

Related: Amidst mass industry layoffs, Konami is raising salaries for the fourth year in a row – “Human resources are the source of our competitiveness”

PlayStation creator Ken Kutaragi suggests the video game industry hasn’t caught up with the evolution of AI. “It’s like there’s no actual progress being made”

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

Generative AI platforms share a few traits with consultants. Both are now ubiquitous, both seek to embed themselves in business operations and both ingest valuable information from elsewhere and sell it back to users as tailored solutions to specific questions.

It remains a stretch to imagine AI replacing consultants completely, as some doomsters predict and plenty of consulting’s critics hope. But it is also clear that wider access to plausible, well-informed, deftly presented advice, supplied by a machine rather than an expensive PowerPoint-toting human being, is posing new challenges to the business model.

At Accenture’s investor day this week, chief executive Julie Sweet mounted the case for the defence. It is based on her conviction that big IT-centric consultancies like hers have “the breadth and the talent to be able to help [clients] achieve their [AI] ambitions”. Meanwhile, strategy consultants such as McKinsey, Boston Consulting Group, Bain and countless smaller boutiques stand ready to offer the road maps that companies need.

The big firms were the superintelligence once, or claimed they were. But AI could undermine their old power “to persuade people that [the] ability to solve problems was just as inspired as the solutions themselves”, to quote Duff McDonald’s 2013 history of McKinsey, The Firm. 

Clients are increasingly looking to machines to do the problem solving, particularly at the less complex end of the market. Where companies require help with the most difficult AI problems and adoption, they may be more likely to turn to consultancies that can attract the most skilled deployment engineers. Indeed, AI labs and talent magnets such as Anthropic and OpenAI are themselves now chasing high-end AI consulting business. 

The economics of the established consultants are creaking as a result. Simple mandates used to be overpriced and farmed out to junior staff, subsidising the trickier work. Companies are now taking the basic jobs in-house and handing them to AI to complete. Consultants’ inability to bill the hours spent prepping a slide deck that an AI agent can churn out easily should eventually focus clients and their advisers more sharply on payment for results. That would be positive. But the transition could be painful.

Still, consultancies have not survived as long as they have by missing disruptive business trends. They spotted early that AI would be the next technology both to discombobulate and excite the boardroom — and leaned into it. Many firms have already hedged their bets by establishing cheaper automated offerings that clients can purchase off the shelf. They have emphasised the juicy efficiencies AI offers and embraced the technology quickly in their own work, demonstrating its value (and sometimes inadvertently highlighting its pitfalls by, for instance, publishing reports polluted by AI hallucinations). 

Change is usually good news for change experts. From the arrival of the internet, via digitisation, Big Data and the cloud, consultancies have found ways to convince clients to pay for advice, reassurance and technical input. Sometimes they win mandates because boards lack the guts to push through destabilising changes on their own and prefer to pay “a few hooded executioners”, to quote Alexander Starritt’s description of a fictionalised McKinsey restructuring team in his state-of-capitalism novel Drayton and Mackenzie. Sometimes executives need someone to blame if their strategy stumbles.

More positively, directors and executives will often reach for a firm hand to hold when their world is shaken by uncertainty. As AI becomes ever more present and capable, the question is not whether businesses will still need that reassurance, but how much of it they will want delivered by flesh-and-blood rather than silicon. Consultants will have to polish their pitch once more.

The staggering rise of AI has become a political flashpoint, especially when it comes to the data centers needed to train and operate it.

As the industry booms, many people are increasingly worried about the environmental impacts of these hulking buildings, from their seemingly insatiable demand for water and energy, to the heat and noise they produce.

Nearly 90% of Americans say they’re more concerned than excited about the rise of AI data centers, according to a CNN poll conducted by SSRS in September.

Information and statistics are flying around, with some groups framing the AI data center boom as a potential environmental catastrophe, while others say concerns are overblown or even fake. It’s adding up to a confusing picture.

So, just how bad is AI for the environment?

Well, to be annoying, it’s complicated. The impacts of each data center depend on a slew of factors, including where it’s located, the kind of electricity that powers it and the technology it uses.

There is also a dearth of fully transparent and detailed data from AI providers, making it tricky to build an accurate picture of impacts. Instead, we rely on estimates.

To make the whole picture even more challenging, it’s difficult to project future impacts because there is so little certainty about how AI technology and demand will develop.

But here’s what we can say at the moment:

It has big energy demands…

Training and running AI models require large amounts of computing power, and that takes a ton of energy.

AI’s energy efficiency is improving at a fast clip, but more people are using the technology and there’s been a boom in particularly energy-intensive forms.

Simple text queries, for example, typically use less electricity than a TV would over the same time period, according an analysis by the International Energy Agency. But, more energy-intensive applications, such as video generation, sometimes require thousands of times more energy. The electricity consumption of AI data centers surged by 50% last year, the IEA found.

Around 0.5% of the world’s electricity was consumed by AI data centers in 2025, according to an analysis from Our World in Data, which projects that figure will reach 1.5% by 2030.

It might sound small, but data centers are very geographically concentrated. Nearly half of their global electricity consumption happens in the United States, where data centers could account for up to 12% of electricity consumption by 2028 as AI rapidly expands, according to an analysis from Lawrence Berkeley National Laboratory.

… and that means climate (and health) impacts

Because the world’s electricity is still dominated by fossil fuels, it adds up to more planet-heating pollution. There are big concerns that the AI boom is pushing climate targets even further out of reach.

Where a data center gets its power from is key; the more clean energy is on the grid, the lower its climate impact. Currently, global data centers are powered mostly by coal (30%), followed by renewables at 27% and natural gas at 26%.

As AI electricity demand soars, some is being met with clean energy, but there are also plenty of planet-warming fossil fuels in the mix.

Natural gas is emerging as a key data center power source in the US. The country has the most gas-fired power capacity in development globally, with more than a third intended to power AI, according to a January report from Global Energy Monitor.

Data centers are increasingly building their own energy infrastructure to avoid straining local grids and driving up people’s bills. This often takes the form of gas turbines, which are less efficient than big power plants and can not only bring hefty climate impacts but also health costs, due to harmful air pollution, including nitrogen oxides and fine particulate matter.

The public health burden of US data centers could reach $20 billion by 2028, according to one recent study.

There’s an oft-cited statistic that a 100-word email generated by an AI chatbot requires about one bottle of water — just over 2 cups — to produce.

But there’s another estimate, from Sam Altman, the CEO of OpenAI, who says the average ChatGPT query uses about one fifteenth of a teaspoon of water.

So which is right? As with many data center impacts, it depends on context and on what people choose to measure.

AI data centers use water directly to cool their equipment, but also indirectly to create the power on which they run.

The water impacts of an individual data center will depend on its specific circumstances: the technology it uses, the kind of energy it relies on and the local water situation.

Some regions, for example, have plentiful water resources; others face high stress. Around two-thirds of data centers built or in construction in the US since 2022 are in areas of water stress, according to a 2025 Bloomberg analysis.

Tech companies say they are introducing measures to slash water use, including by developing new cooling technologies. Some analysts say water concerns have been vastly overestimated and that the burger you eat or cotton shirt you wear require much more water.

But AI’s water impacts are far from minimal. US data centers could consume up to 110 billion gallons a year by 2030, equivalent to 7% of total annual consumptive use in US public water systems, according to a March analysis by scientists from the University of California, Riverside and other institutions.

Indirect water use for electricity could push this figure much higher. A recent report from Ceres, a nonprofit advocacy organization, analyzed data centers across seven states and found they used around 3.4 trillion gallons of freshwater a year for electricity generation, equivalent to roughly 12 times the annual water use of Los Angeles, Phoenix and Washington, DC, combined.

Some of the key concerns people have about data centers revolve around quality of life impacts — including noise pollution.

Data centers can produce a cacophony of noises: the low thrum of cooling systems, the drone of gas turbines or diesel generators and the humming of equipment.

Some communities have filed noise pollution lawsuits, including residents near a huge data center in Vineland, New Jersey.

There are steps data centers can take to reduce their impact, including soundproofing equipment and the building itself. Some experts are urging state and local governments to beef up noise regulations.

There are also fears the data centers could create “heat islands,” warming the air and land around them — a phenomenon that would add to the temperature rises already happening due to fossil fuel pollution.

A study by scientists at Arizona State University found data center heat pollution in Phoenix warmed downwind neighborhoods by as much as 4 degrees Fahrenheit.

“Even if these data centers only contribute to an additional heat island magnitude of 1 degree or 2 degrees, that can still have a very significant impact on our lives,” said lead author David Sailor, director of ASU’s School of Geographical Sciences and Urban Planning.

More research is needed, however, to really pin down these impacts, and some experts argue there is little difference between a data center and any other industrial facility when it comes to heat.

One of the more overlooked aspects of AI are the minerals and metals it requires, including for semiconductors and data storage components.

Mining for these raw materials, often in the Global South, has been linked to air and water pollution, deforestation, human health risks and climate pollution.

And as AI scales and advances, it will generate electronic waste, a rapidly-growing waste stream that is toxic if not disposed of properly. Currently only about a fifth of e-waste is recycled.

There’s also concern that AI’s demand for these minerals put it in direct competition with sectors that are critical for reducing planet heating pollution, namely wind and solar energy.

Experts say AI’s environmental harms are not locked in. Instead, its impacts will depend on the choices we make: Where data centers are located, how they are powered, how they are regulated, how transparent AI providers must be about impacts along their full supply chains.

Before a new AI data center is built, it needs to be clear “who benefits from it,” said Kaveh Madani, director of the United Nations University Institute for Water, Environment and Health. That includes asking questions like: Does it bring jobs? Will it improve the local economy? Will it take water from local users who need it more, like farmers?

“There are trade-offs like any other technology,” Madani said. “It has has downsides that need to be regulated, monitored, and mitigated.”

Powerful AI models are going on hacking sprees — or helping humans with them — left and right. If the AI industry doesn’t get its collective act together and stop its products from being used to such dangerous ends, it could be only a matter of time before one of these AI-powered hacks targets our essential infrastructure.

This is a scenario that some AI researchers are apparently already preparing for. Andreas Stuhlmüller, CEO and cofounder of the AI research company Elicit, recently rattled off a prepper-esque checklist of how to prepare for when AI-powered cyberattacks could take out financial systems, power grids, the internet, and all the other infrastructure that underpins modern life.

“Cash: Get a few weeks of spending in cash,” he recommended in a tweet. Then he suggests, in greater detail than we’ll get into here, getting a Starlink terminal in case your local internet network goes down; a battery, generator and solar panels for power; weeks worth of water and food; and a hand-crank emergency radio.

Stuhlmüller’s paranoid recommendations come after a shocking cyberattack that targeted the refrigerator systems used in hundreds of grocery stores across France. The as-yet unknown perpetrators cranked up the fridge temperatures up to 122 degrees Fahrenheit, spoiling food and even causing fires. The CEO of one chain targeted by the attack, Intermarche, said it caused losses amounting to “tens ​of millions of euros nationwide in a single day.”

AI’s role in the attack hasn’t been confirmed. But given AI’s ubiquity among coders, it’s likely that it was used to some degree. On top of churning out code, the tech is alarmingly good at identifying vulnerabilities. A recent CrowdStrike report found that there was a nearly 90 percent year over year increase in AI-enabled cyberattacks.

There’s also, of course, the recent disclosures from frontier AI labs like OpenAI, Anthropic, and Meta, over how their agentic AI systems broke out of their testing environments and launched cyberattacks against other companies’ websites and computer systems.

These incidents represent an alarming escalation in the threat AI poses to cybersecurity systems, not only assisting hacks but carrying themselves autonomously while also avoiding the detection of their human overseers.

All that being said, it’s a major leap to say that AI models — or humans using them — will go from hacking websites and supermarket fridges to your local power grid and banking system. But we can hardly blame anyone for being worried about this possible eventuality. Better safe than sorry, right?

More on AI: Godfathers of AI Say Models Are So Out of Control That They Need to Be Airgapped for the Safety of Us All

Story Continues

Elorian, the Palo Alto AI startup that raised a $55 million seed round at a $300 million valuation before it launched a product, released its first model on Friday, TechCrunch has exclusively learned.

The model, Glimpse, is launching as a research preview with API access and a chat interface. The thesis? Today’s AI models, for all their skill with language and code, still can’t really see. 

Most multimodal models look at an image once, convert it to tokens, and then reason about it in text — like describing a photo from memory. Glimpse keeps going back to the source. Mid-reasoning, it can crop into a region, pull specific frames from a video, or run code to measure or compare what it’s looking at, a mechanism Elorian calls a “visual scratch space.” Think of someone pinching to zoom on a phone photo to double-check what they thought they saw.

For former Google DeepMind researcher and Elorian CEO Andrew Dai, one of the clearest signs that AI still couldn’t really see came on game night. 

Last year he photographed a board game he’d just finished with friends and asked a frontier model who was winning. It couldn’t say. Board games “each have their own visual language, each have their own iconography, symbols,” he told TechCrunch. “To an adult, it’s quite easy to understand what’s going on with minimal text, but these models aren’t able to intuit what the symbols mean.”

“Humans can look at Ikea diagrams, with no text, and assemble Ikea furniture, right? Why can’t AI models?” he said. 

The approach isn’t completely new — OpenAI added crop-and-zoom during reasoning to its o-series models in 2025 — but Elorian’s pitch is that it built the whole model around it, cheaply. The company says a small team trained Glimpse on a few hundred GPUs and that the model is substantially smaller than the frontier systems it’s benchmarked against. 

Glimpse costs $0.60 per million tokens in and $3 per million tokens out — cheaper than mid-level models like Sonnet and Sol but more expensive than lightweight models like Haiku and Luna.

But according to Elorian’s evaluations, Glimpse outperforms even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks. 

Elorian came out of stealth six months ago with a round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.

The initial targets are companies sitting on lots of video: robotics firms, scientific labs, media, and creative companies. Longer term, Elorian wants to move from seeing to doing. That means building a model of a scene using that to choose an action, including directly controlling robots. Next year, Dai said, the company aims to build “the world’s best visual reasoning model,” one that leads on every visual reasoning test. 

This page either does not exist or is currently unavailable.

From here you can either hit the “back” button on your browser to return to the previous page, or visit the ABCNews.com Home Page. You can also search for something on our site below.

STATUS CODE: 404

Benchmarks are the rulers of AI, in both senses of the word. As measuring sticks, they work according to an extremely simple logic: If you want to know what an AI model can do, devise a standardized test to gauge it. But AI benchmarks can also dominate the behavior of the entire field.

In 1987, the Defense Advanced Research Projects Agency’s first benchmark on automatic speech-to-text translation showed that state-of-the-art AI could not outperform so-called “dumb” statistical approaches. The poor performance helped to usher in an AI winter. For 25 years, interest in AI research waned and funding dried up.

Then, in 2012, another benchmark called ImageNet swung the pendulum back. An eight-layer neural network trounced other algorithms at classifying images and kicked off a “deep learning revolution.”

Nowadays, new AI benchmarks, geared at large language models such as OpenAI’s GPT-6 Astra, get created and solved with clockwork regularity, purporting to measure everything from domain-specific expertise to “general intelligence.” But rather than hail this trend as proof of AI’s skyrocketing potential, many in the field — skeptics and boosters alike — take it as evidence that benchmarking itself has become fatally flawed.

Moritz Hardt, a computer scientist and a director at the Max Planck Institute for Intelligent Systems, agrees with this diagnosis. Shortly after the ImageNet result, he began to think that machine learning systems could game leaderboards “without doing anything intelligent.” “There’s no two ways about it,” he said. “The benchmarking ecosystem is broken.”

In his 2026 book The Emerging Science of Machine Learning Benchmarks, Hardt makes the provocative case that these measuring sticks have been broken for decades already: Benchmarks, he writes, are “the mistake that made machine learning.” Notably, he argues, instead of creating a replicability crisis, they drove legitimate progress anyway — “mostly by sheer luck,” in his view. By understanding why, we can explain how benchmarks have since lost this accidental credibility in the age of LLMs, and how they might regain it.

Quanta spoke to Hardt about what went wrong, and right, with benchmarks, and where to take them next. The interview has been condensed and edited for clarity.

AI agents from multiple providers accessed personal data and confidential material at Indian organizations during research and testing, India’s cybersecurity agency said Thursday.

The Indian Computer Emergency Response Team (CERT-In) said it had received “multiple voluntary disclosures from AI model providers” whose models, during research and evaluation runs on the open internet, “had reached the live systems of Indian organizations.”

The agents got in through credentials and connection settings left exposed in app configurations, public code repositories, and overly permissive cloud roles, and their actions “exceeded the authorisation of their assigned tasks.”

This makes India the latest stop on what has become a world tour for rogue AI agents. In July, OpenAI’s agents escaped a test environment and hacked French company Hugging Face — kicking off a summer of rogue AI agent activity. Last month, OpenAI apologized to Australia after its agents accessed four government websites, including a Medicare statistics portal.

At Hugging Face, OpenAI’s agents started the same way but then found and exploited software flaws to run code on the company’s servers. CERT-In described no such hacking in India, where the agents appear to have simply used the access left lying around.

Still, the Indian cases appear to be worse in one respect, which is what the agents saw: In Australia, OpenAI said it found no evidence its agents touched individuals’ medical or criminal records, even after they wrote files inside a Medicare statistics system, while in India, agents read personal data and confidential business records. And unlike earlier incidents, each disclosed by a single lab, India’s case involved agents from multiple providers.

CERT-In did not name the providers or affected organizations, or say whether the agents altered any systems. TechCrunch has reached out to CERT-In for more information.

For now, the closest parallel seems to be Anthropic’s July disclosure of Claude breaking into three companies’ systems during testing using basic techniques like weak passwords.

In India, as in Anthropic’s cases, the affected organizations had not detected the intrusions on their own. The disclosures have left governments relying on AI companies causing these breaches to voluntarily disclose rogue AI agent activity.