How to build a better AI benchmark

Stay Ahead, Stay ONMINE

How to build a better AI benchmark

It’s not easy being one of Silicon Valley’s favorite benchmarks. SWE-Bench (pronounced “swee bench”) launched in November 2024 to evaluate an AI model’s coding skill, using more than 2,000 real-world programming problems pulled from the public GitHub repositories of 12 different Python-based projects. In the months since then, it’s quickly become one of the most popular tests in AI. A SWE-Bench score has become a mainstay of major model releases from OpenAI, Anthropic, and Google—and outside of foundation models, the fine-tuners at AI firms are in constant competition to see who can rise above the pack. The top of the leaderboard is a pileup between three different fine tunings of Anthropic’s Claude Sonnet model and Amazon’s Q developer agent. Auto Code Rover—one of the Claude modifications—nabbed the number two spot in November, and was acquired just three months later. Despite all the fervor, this isn’t exactly a truthful assessment of which model is “better.” As the benchmark has gained prominence, “you start to see that people really want that top spot,” says John Yang, a researcher on the team that developed SWE-Bench at Princeton University. As a result, entrants have begun to game the system—which is pushing many others to wonder whether there’s a better way to actually measure AI achievement. Developers of these coding agents aren’t necessarily doing anything as straightforward cheating, but they’re crafting approaches that are too neatly tailored to the specifics of the benchmark. The initial SWE-Bench test set was limited to programs written in Python, which meant developers could gain an advantage by training their models exclusively on Python code. Soon, Yang noticed that high-scoring models would fail completely when tested on different programming languages—revealing an approach to the test that he describes as “gilded.” “It looks nice and shiny at first glance, but then you try to run it on a different language and the whole thing just kind of falls apart,” Yang says. “At that point, you’re not designing a software engineering agent. You’re designing to make a SWE-Bench agent, which is much less interesting.” The SWE-Bench issue is a symptom of a more sweeping—and complicated—problem in AI evaluation, and one that’s increasingly sparking heated debate: The benchmarks the industry uses to guide development are drifting further and further away from evaluating actual capabilities, calling their basic value into question. Making the situation worse, several benchmarks, most notably FrontierMath and Chatbot Arena, have recently come under heat for an alleged lack of transparency. Nevertheless, benchmarks still play a central role in model development, even if few experts are willing to take their results at face value. OpenAI cofounder Andrej Karpathy recently described the situation as “an evaluation crisis”: the industry has fewer trusted methods for measuring capabilities and no clear path to better ones. “Historically, benchmarks were the way we evaluated AI systems,” says Vanessa Parli, director of research at Stanford University’s Institute for Human-Centered AI. “Is that the way we want to evaluate systems going forward? And if it’s not, what is the way?” A growing group of academics and AI researchers are making the case that the answer is to go smaller, trading sweeping ambition for an approach inspired by the social sciences. Specifically, they want to focus more on testing validity, which for quantitative social scientists refers to how well a given questionnaire measures what it’s claiming to measure—and, more fundamentally, whether what it is measuring has a coherent definition. That could cause trouble for benchmarks assessing hazily defined concepts like “reasoning” or “scientific knowledge”—and for developers aiming to reach the much-hyped goal of artificial general intelligence—but it would put the industry on firmer ground as it looks to prove the worth of individual models. “Taking validity seriously means asking folks in academia, industry, or wherever to show that their system does what they say it does,” says Abigail Jacobs, a University of Michigan professor who is a central figure in the new push for validity. “I think it points to a weakness in the AI world if they want to back off from showing that they can support their claim.” The limits of traditional testing If AI companies have been slow to respond to the growing failure of benchmarks, it’s partially because the test-scoring approach has been so effective for so long. One of the biggest early successes of contemporary AI was the ImageNet challenge, a kind of antecedent to contemporary benchmarks. Released in 2010 as an open challenge to researchers, the database held more than 3 million images for AI systems to categorize into 1,000 different classes. Crucially, the test was completely agnostic to methods, and any successful algorithm quickly gained credibility regardless of how it worked. When an algorithm called AlexNet broke through in 2012, with a then unconventional form of GPU training, it became one of the foundational results of modern AI. Few would have guessed in advance that AlexNet’s convolutional neural nets would be the secret to unlocking image recognition—but after it scored well, no one dared dispute it. (One of AlexNet’s developers, Ilya Sutskever, would go on to cofound OpenAI.) A large part of what made this challenge so effective was that there was little practical difference between ImageNet’s object classification challenge and the actual process of asking a computer to recognize an image. Even if there were disputes about methods, no one doubted that the highest-scoring model would have an advantage when deployed in an actual image recognition system. But in the 12 years since, AI researchers have applied that same method-agnostic approach to increasingly general tasks. SWE-Bench is commonly used as a proxy for broader coding ability, while other exam-style benchmarks often stand in for reasoning ability. That broad scope makes it difficult to be rigorous about what a specific benchmark measures—which, in turn, makes it hard to use the findings responsibly. Where things break down Anka Reuel, a PhD student who has been focusing on the benchmark problem as part of her research at Stanford, has become convinced the evaluation problem is the result of this push toward generality. “We’ve moved from task-specific models to general-purpose models,” Reuel says. “It’s not about a single task anymore but a whole bunch of tasks, so evaluation becomes harder.” Like the University of Michigan’s Jacobs, Reuel thinks “the main issue with benchmarks is validity, even more than the practical implementation,” noting: “That’s where a lot of things break down.” For a task as complicated as coding, for instance, it’s nearly impossible to incorporate every possible scenario into your problem set. As a result, it’s hard to gauge whether a model is scoring better because it’s more skilled at coding or because it has more effectively manipulated the problem set. And with so much pressure on developers to achieve record scores, shortcuts are hard to resist. For developers, the hope is that success on lots of specific benchmarks will add up to a generally capable model. But the techniques of agentic AI mean a single AI system can encompass a complex array of different models, making it hard to evaluate whether improvement on a specific task will lead to generalization. “There’s just many more knobs you can turn,” says Sayash Kapoor, a computer scientist at Princeton and a prominent critic of sloppy practices in the AI industry. “When it comes to agents, they have sort of given up on the best practices for evaluation.” In a paper from last July, Kapoor called out specific issues in how AI models were approaching the WebArena benchmark, designed by Carnegie Mellon University researchers in 2024 as a test of an AI agent’s ability to traverse the web. The benchmark consists of more than 800 tasks to be performed on a set of cloned websites mimicking Reddit, Wikipedia, and others. Kapoor and his team identified an apparent hack in the winning model, called STeP. STeP included specific instructions about how Reddit structures URLs, allowing STeP models to jump directly to a given user’s profile page (a frequent element of WebArena tasks). This shortcut wasn’t exactly cheating, but Kapoor sees it as “a serious misrepresentation of how well the agent would work had it seen the tasks in WebArena for the first time.” Because the technique was successful, though, a similar policy has since been adopted by OpenAI’s web agent Operator. (“Our evaluation setting is designed to assess how well an agent can solve tasks given some instruction about website structures and task execution,” an OpenAI representative said when reached for comment. “This approach is consistent with how others have used and reported results with WebArena.” STeP did not respond to a request for comment.) Further highlighting the problem with AI benchmarks, late last month Kapoor and a team of researchers wrote a paper that revealed significant problems in Chatbot Arena, the popular crowdsourced evaluation system. According to the paper, the leaderboard was being manipulated; many top foundation models were conducting undisclosed private testing and releasing their scores selectively. Today, even ImageNet itself, the mother of all benchmarks, has started to fall victim to validity problems. A 2023 study from researchers at the University of Washington and Google Research found that when ImageNet-winning algorithms were pitted against six real-world data sets, the architecture improvement “resulted in little to no progress,” suggesting that the external validity of the test had reached its limit. Going smaller For those who believe the main problem is validity, the best fix is reconnecting benchmarks to specific tasks. As Reuel puts it, AI developers “have to resort to these high-level benchmarks that are almost meaningless for downstream consumers, because the benchmark developers can’t anticipate the downstream task anymore.” So what if there was a way to help the downstream consumers identify this gap? In November 2024, Reuel launched a public ranking project called BetterBench, which rates benchmarks on dozens of different criteria, such as whether the code has been publicly documented. But validity is a central theme, with particular criteria challenging designers to spell out what capability their benchmark is testing and how it relates to the tasks that make up the benchmark. “You need to have a structural breakdown of the capabilities,” Reuel says. “What are the actual skills you care about, and how do you operationalize them into something we can measure?” The results are surprising. One of the highest-scoring benchmarks is also the oldest: the Arcade Learning Environment (ALE), established in 2013 as a way to test models’ ability to learn how to play a library of Atari 2600 games. One of the lowest-scoring is the Massive Multitask Language Understanding (MMLU) benchmark, a widely used test for general language skills; by the standards of BetterBench, the connection between the questions and the underlying skill was too poorly defined. BetterBench hasn’t meant much for the reputations of specific benchmarks, at least not yet; MMLU is still widely used, and ALE is still marginal. But the project has succeeded in pushing validity into the broader conversation about how to fix benchmarks. In April, Reuel quietly joined a new research group hosted by Hugging Face, the University of Edinburgh, and EleutherAI, where she’ll develop her ideas on validity and AI model evaluation with other figures in the field. (An official announcement is expected later this month.) Irene Solaiman, Hugging Face’s head of global policy, says the group will focus on building valid benchmarks that go beyond measuring straightforward capabilities. “There’s just so much hunger for a good benchmark off the shelf that already works,” Solaiman says. “A lot of evaluations are trying to do too much.” Increasingly, the rest of the industry seems to agree. In a paper in March, researchers from Google, Microsoft, Anthropic, and others laid out a new framework for improving evaluations—with validity as the first step. “AI evaluation science must,” the researchers argue, “move beyond coarse grained claims of ‘general intelligence’ towards more task-specific and real-world relevant measures of progress.” Measuring the “squishy” things To help make this shift, some researchers are looking to the tools of social science. A February position paper argued that “evaluating GenAI systems is a social science measurement challenge,” specifically unpacking how the validity systems used in social measurements can be applied to AI benchmarking. The authors, largely employed by Microsoft’s research branch but joined by academics from Stanford and the University of Michigan, point to the standards that social scientists use to measure contested concepts like ideology, democracy, and media bias. Applied to AI benchmarks, those same procedures could offer a way to measure concepts like “reasoning” and “math proficiency” without slipping into hazy generalizations. In the social science literature, it’s particularly important that metrics begin with a rigorous definition of the concept measured by the test. For instance, if the test is to measure how democratic a society is, it first needs to establish a definition for a “democratic society” and then establish questions that are relevant to that definition. To apply this to a benchmark like SWE-Bench, designers would need to set aside the classic machine learning approach, which is to collect programming problems from GitHub and create a scheme to validate answers as true or false. Instead, they’d first need to define what the benchmark aims to measure (“ability to resolve flagged issues in software,” for instance), break that into subskills (different types of problems or types of program that the AI model can successfully process), and then finally assemble questions that accurately cover the different subskills. It’s a profound change from how AI researchers typically approach benchmarking—but for researchers like Jacobs, a coauthor on the February paper, that’s the whole point. “There’s a mismatch between what’s happening in the tech industry and these tools from social science,” she says. “We have decades and decades of thinking about how we want to measure these squishy things about humans.” Even though the idea has made a real impact in the research world, it’s been slow to influence the way AI companies are actually using benchmarks. The last two months have seen new model releases from OpenAI, Anthropic, Google, and Meta, and all of them lean heavily on multiple-choice knowledge benchmarks like MMLU—the exact approach that validity researchers are trying to move past. After all, model releases are, for the most part, still about showing increases in general intelligence, and broad benchmarks continue to be used to back up those claims. For some observers, that’s good enough. Benchmarks, Wharton professor Ethan Mollick says, are “bad measures of things, but also they’re what we’ve got.” He adds: “At the same time, the models are getting better. A lot of sins are forgiven by fast progress.” For now, the industry’s long-standing focus on artificial general intelligence seems to be crowding out a more focused validity-based approach. As long as AI models can keep growing in general intelligence, then specific applications don’t seem as compelling—even if that leaves practitioners relying on tools they no longer fully trust. “This is the tightrope we’re walking,” says Hugging Face’s Solaiman. “It’s too easy to throw the system out, but evaluations are really helpful in understanding our models, even with these limitations.” Russell Brandom is a freelance writer covering artificial intelligence. He lives in Brooklyn with his wife and two cats. This story was supported by a grant from the Tarbell Center for AI Journalism.

It’s not easy being one of Silicon Valley’s favorite benchmarks.

SWE-Bench (pronounced “swee bench”) launched in November 2024 to evaluate an AI model’s coding skill, using more than 2,000 real-world programming problems pulled from the public GitHub repositories of 12 different Python-based projects.

In the months since then, it’s quickly become one of the most popular tests in AI. A SWE-Bench score has become a mainstay of major model releases from OpenAI, Anthropic, and Google—and outside of foundation models, the fine-tuners at AI firms are in constant competition to see who can rise above the pack. The top of the leaderboard is a pileup between three different fine tunings of Anthropic’s Claude Sonnet model and Amazon’s Q developer agent. Auto Code Rover—one of the Claude modifications—nabbed the number two spot in November, and was acquired just three months later.

Despite all the fervor, this isn’t exactly a truthful assessment of which model is “better.” As the benchmark has gained prominence, “you start to see that people really want that top spot,” says John Yang, a researcher on the team that developed SWE-Bench at Princeton University. As a result, entrants have begun to game the system—which is pushing many others to wonder whether there’s a better way to actually measure AI achievement.

Developers of these coding agents aren’t necessarily doing anything as straightforward cheating, but they’re crafting approaches that are too neatly tailored to the specifics of the benchmark. The initial SWE-Bench test set was limited to programs written in Python, which meant developers could gain an advantage by training their models exclusively on Python code. Soon, Yang noticed that high-scoring models would fail completely when tested on different programming languages—revealing an approach to the test that he describes as “gilded.”

“It looks nice and shiny at first glance, but then you try to run it on a different language and the whole thing just kind of falls apart,” Yang says. “At that point, you’re not designing a software engineering agent. You’re designing to make a SWE-Bench agent, which is much less interesting.”

The SWE-Bench issue is a symptom of a more sweeping—and complicated—problem in AI evaluation, and one that’s increasingly sparking heated debate: The benchmarks the industry uses to guide development are drifting further and further away from evaluating actual capabilities, calling their basic value into question. Making the situation worse, several benchmarks, most notably FrontierMath and Chatbot Arena, have recently come under heat for an alleged lack of transparency. Nevertheless, benchmarks still play a central role in model development, even if few experts are willing to take their results at face value. OpenAI cofounder Andrej Karpathy recently described the situation as “an evaluation crisis”: the industry has fewer trusted methods for measuring capabilities and no clear path to better ones.

“Historically, benchmarks were the way we evaluated AI systems,” says Vanessa Parli, director of research at Stanford University’s Institute for Human-Centered AI. “Is that the way we want to evaluate systems going forward? And if it’s not, what is the way?”

A growing group of academics and AI researchers are making the case that the answer is to go smaller, trading sweeping ambition for an approach inspired by the social sciences. Specifically, they want to focus more on testing validity, which for quantitative social scientists refers to how well a given questionnaire measures what it’s claiming to measure—and, more fundamentally, whether what it is measuring has a coherent definition. That could cause trouble for benchmarks assessing hazily defined concepts like “reasoning” or “scientific knowledge”—and for developers aiming to reach the much–hyped goal of artificial general intelligence—but it would put the industry on firmer ground as it looks to prove the worth of individual models.

“Taking validity seriously means asking folks in academia, industry, or wherever to show that their system does what they say it does,” says Abigail Jacobs, a University of Michigan professor who is a central figure in the new push for validity. “I think it points to a weakness in the AI world if they want to back off from showing that they can support their claim.”

The limits of traditional testing

If AI companies have been slow to respond to the growing failure of benchmarks, it’s partially because the test-scoring approach has been so effective for so long.

One of the biggest early successes of contemporary AI was the ImageNet challenge, a kind of antecedent to contemporary benchmarks. Released in 2010 as an open challenge to researchers, the database held more than 3 million images for AI systems to categorize into 1,000 different classes.

Crucially, the test was completely agnostic to methods, and any successful algorithm quickly gained credibility regardless of how it worked. When an algorithm called AlexNet broke through in 2012, with a then unconventional form of GPU training, it became one of the foundational results of modern AI. Few would have guessed in advance that AlexNet’s convolutional neural nets would be the secret to unlocking image recognition—but after it scored well, no one dared dispute it. (One of AlexNet’s developers, Ilya Sutskever, would go on to cofound OpenAI.)

A large part of what made this challenge so effective was that there was little practical difference between ImageNet’s object classification challenge and the actual process of asking a computer to recognize an image. Even if there were disputes about methods, no one doubted that the highest-scoring model would have an advantage when deployed in an actual image recognition system.

But in the 12 years since, AI researchers have applied that same method-agnostic approach to increasingly general tasks. SWE-Bench is commonly used as a proxy for broader coding ability, while other exam-style benchmarks often stand in for reasoning ability. That broad scope makes it difficult to be rigorous about what a specific benchmark measures—which, in turn, makes it hard to use the findings responsibly.

Where things break down

Anka Reuel, a PhD student who has been focusing on the benchmark problem as part of her research at Stanford, has become convinced the evaluation problem is the result of this push toward generality. “We’ve moved from task-specific models to general-purpose models,” Reuel says. “It’s not about a single task anymore but a whole bunch of tasks, so evaluation becomes harder.”

Like the University of Michigan’s Jacobs, Reuel thinks “the main issue with benchmarks is validity, even more than the practical implementation,” noting: “That’s where a lot of things break down.” For a task as complicated as coding, for instance, it’s nearly impossible to incorporate every possible scenario into your problem set. As a result, it’s hard to gauge whether a model is scoring better because it’s more skilled at coding or because it has more effectively manipulated the problem set. And with so much pressure on developers to achieve record scores, shortcuts are hard to resist.

For developers, the hope is that success on lots of specific benchmarks will add up to a generally capable model. But the techniques of agentic AI mean a single AI system can encompass a complex array of different models, making it hard to evaluate whether improvement on a specific task will lead to generalization. “There’s just many more knobs you can turn,” says Sayash Kapoor, a computer scientist at Princeton and a prominent critic of sloppy practices in the AI industry. “When it comes to agents, they have sort of given up on the best practices for evaluation.”

In a paper from last July, Kapoor called out specific issues in how AI models were approaching the WebArena benchmark, designed by Carnegie Mellon University researchers in 2024 as a test of an AI agent’s ability to traverse the web. The benchmark consists of more than 800 tasks to be performed on a set of cloned websites mimicking Reddit, Wikipedia, and others. Kapoor and his team identified an apparent hack in the winning model, called STeP. STeP included specific instructions about how Reddit structures URLs, allowing STeP models to jump directly to a given user’s profile page (a frequent element of WebArena tasks).

This shortcut wasn’t exactly cheating, but Kapoor sees it as “a serious misrepresentation of how well the agent would work had it seen the tasks in WebArena for the first time.” Because the technique was successful, though, a similar policy has since been adopted by OpenAI’s web agent Operator. (“Our evaluation setting is designed to assess how well an agent can solve tasks given some instruction about website structures and task execution,” an OpenAI representative said when reached for comment. “This approach is consistent with how others have used and reported results with WebArena.” STeP did not respond to a request for comment.)

Further highlighting the problem with AI benchmarks, late last month Kapoor and a team of researchers wrote a paper that revealed significant problems in Chatbot Arena, the popular crowdsourced evaluation system. According to the paper, the leaderboard was being manipulated; many top foundation models were conducting undisclosed private testing and releasing their scores selectively.

Today, even ImageNet itself, the mother of all benchmarks, has started to fall victim to validity problems. A 2023 study from researchers at the University of Washington and Google Research found that when ImageNet-winning algorithms were pitted against six real-world data sets, the architecture improvement “resulted in little to no progress,” suggesting that the external validity of the test had reached its limit.

Going smaller

For those who believe the main problem is validity, the best fix is reconnecting benchmarks to specific tasks. As Reuel puts it, AI developers “have to resort to these high-level benchmarks that are almost meaningless for downstream consumers, because the benchmark developers can’t anticipate the downstream task anymore.” So what if there was a way to help the downstream consumers identify this gap?

In November 2024, Reuel launched a public ranking project called BetterBench, which rates benchmarks on dozens of different criteria, such as whether the code has been publicly documented. But validity is a central theme, with particular criteria challenging designers to spell out what capability their benchmark is testing and how it relates to the tasks that make up the benchmark.

“You need to have a structural breakdown of the capabilities,” Reuel says. “What are the actual skills you care about, and how do you operationalize them into something we can measure?”

The results are surprising. One of the highest-scoring benchmarks is also the oldest: the Arcade Learning Environment (ALE), established in 2013 as a way to test models’ ability to learn how to play a library of Atari 2600 games. One of the lowest-scoring is the Massive Multitask Language Understanding (MMLU) benchmark, a widely used test for general language skills; by the standards of BetterBench, the connection between the questions and the underlying skill was too poorly defined.

BetterBench hasn’t meant much for the reputations of specific benchmarks, at least not yet; MMLU is still widely used, and ALE is still marginal. But the project has succeeded in pushing validity into the broader conversation about how to fix benchmarks. In April, Reuel quietly joined a new research group hosted by Hugging Face, the University of Edinburgh, and EleutherAI, where she’ll develop her ideas on validity and AI model evaluation with other figures in the field. (An official announcement is expected later this month.)

Irene Solaiman, Hugging Face’s head of global policy, says the group will focus on building valid benchmarks that go beyond measuring straightforward capabilities. “There’s just so much hunger for a good benchmark off the shelf that already works,” Solaiman says. “A lot of evaluations are trying to do too much.”

Increasingly, the rest of the industry seems to agree. In a paper in March, researchers from Google, Microsoft, Anthropic, and others laid out a new framework for improving evaluations—with validity as the first step.

“AI evaluation science must,” the researchers argue, “move beyond coarse grained claims of ‘general intelligence’ towards more task-specific and real-world relevant measures of progress.”

Measuring the “squishy” things

To help make this shift, some researchers are looking to the tools of social science. A February position paper argued that “evaluating GenAI systems is a social science measurement challenge,” specifically unpacking how the validity systems used in social measurements can be applied to AI benchmarking.

The authors, largely employed by Microsoft’s research branch but joined by academics from Stanford and the University of Michigan, point to the standards that social scientists use to measure contested concepts like ideology, democracy, and media bias. Applied to AI benchmarks, those same procedures could offer a way to measure concepts like “reasoning” and “math proficiency” without slipping into hazy generalizations.

In the social science literature, it’s particularly important that metrics begin with a rigorous definition of the concept measured by the test. For instance, if the test is to measure how democratic a society is, it first needs to establish a definition for a “democratic society” and then establish questions that are relevant to that definition.

To apply this to a benchmark like SWE-Bench, designers would need to set aside the classic machine learning approach, which is to collect programming problems from GitHub and create a scheme to validate answers as true or false. Instead, they’d first need to define what the benchmark aims to measure (“ability to resolve flagged issues in software,” for instance), break that into subskills (different types of problems or types of program that the AI model can successfully process), and then finally assemble questions that accurately cover the different subskills.

It’s a profound change from how AI researchers typically approach benchmarking—but for researchers like Jacobs, a coauthor on the February paper, that’s the whole point. “There’s a mismatch between what’s happening in the tech industry and these tools from social science,” she says. “We have decades and decades of thinking about how we want to measure these squishy things about humans.”

Even though the idea has made a real impact in the research world, it’s been slow to influence the way AI companies are actually using benchmarks.

The last two months have seen new model releases from OpenAI, Anthropic, Google, and Meta, and all of them lean heavily on multiple-choice knowledge benchmarks like MMLU—the exact approach that validity researchers are trying to move past. After all, model releases are, for the most part, still about showing increases in general intelligence, and broad benchmarks continue to be used to back up those claims.

For some observers, that’s good enough. Benchmarks, Wharton professor Ethan Mollick says, are “bad measures of things, but also they’re what we’ve got.” He adds: “At the same time, the models are getting better. A lot of sins are forgiven by fast progress.”

For now, the industry’s long-standing focus on artificial general intelligence seems to be crowding out a more focused validity-based approach. As long as AI models can keep growing in general intelligence, then specific applications don’t seem as compelling—even if that leaves practitioners relying on tools they no longer fully trust.

“This is the tightrope we’re walking,” says Hugging Face’s Solaiman. “It’s too easy to throw the system out, but evaluations are really helpful in understanding our models, even with these limitations.”

Russell Brandom is a freelance writer covering artificial intelligence. He lives in Brooklyn with his wife and two cats.

This story was supported by a grant from the Tarbell Center for AI Journalism.

Stay Ahead

Explore More Insights

Stay ahead with more perspectives on cutting-edge power, infrastructure, energy, bitcoin and AI solutions. Explore these articles to uncover strategies and insights shaping the future of industries.

Aggressive federal PQE timeline prompts warnings for enterprises

“The U.S. government’s EOs will likely spur accelerated intervention from all major governments and regional political blocs,” the firm said. “CISOs should be prepared for regulations to conflict and contain sovereignty requirements, which will complicate compliance.” Gartner recommends that companies build a PQC inventory and remediation program in 2026 and

Presidential order addresses quantum computing gaps

By comparison, in AI, there are a number of benchmarks comparing AI models on everything from how well they do at math, to their ability to generate images, to whether they can find bugs in computer code. Benchmarks are important so that people can do fair comparisons of quantum systems,

Researchers cast new doubt on Microsoft’s quantum computing advance

In 2018, Microsoft said its researchers had detected evidence of their existence, an apparently major breakthrough it was forced to retract when the data was successfully challenged. Nature’s editors subsequently backed this up with the blunt note: “The results in this manuscript do not represent evidence for the presence of

IBM unveils sub-1 nanometer chip with nearly 100 billion transistors

It’s the world’s first sub-1 nm chip technology, IBM claims. Researcher holds IBM’s sub-1 nm node wafer.

Equinor to invest in additional Troll development to boost European gas supply

Equinor Energy AS and partners will invest more than 4 billion krone ($400 million) in a new subsea development to increase gas production from Troll field in the North Sea. The Troll West Increased gas recovery North (TWIN) expansion—the third step of Troll Phase 3, which produces gas from the Troll West reservoir—could come online as early as 2028, said Gunnar Nakken, Equinor’s senior vice-president for projects and subsea Norway. TWIN is expected to contribute around 11 billion standard cu m of gas. “By simplifying, increasing standardization and reusing existing infrastructure and equipment, we are reducing costs and enabling faster production,” he said. Equinor aims to produce 1.3 million b/d from the Norwegian Continental Shelf (NCS) in 2035 to meet a portion of Europe’s energy needs. Troll field contains about 40% of NCS total gas reserves, with gas from Troll meeting around 10% of Europe’s gas needs. The TWIN project consists of two wells in a template and a pipeline connected to existing subsea infrastructure. The umbilical and MEG line will be extended to the new development. The second step of Troll Phase 3 is expected to come online this year, continuing production from Troll A platform, 80 km northwest of Bergen, Norway, and the Gassco-operated Kollsnes processing plant towards 2030, Equinor said. Equinor is operator of the project with 30.55% interest. Partners are Petoro AS (55.93%), A/S Norske Shell (8.19%), TotalEnergies EP Norge AS (3.69%), and ConocoPhillips Skandinavia AS (1.64%).

Zululand Energy Terminal invites EPC expressions of interest

The proposed 7.5-million tonne/year (tpy) Zululand Energy Terminal (ZET) at the Port of Richards Bay, South Africa, has invited expressions of interest (EOI) from engineering, procurement and construction (EPC) contractors for development of planned LNG regasification infrastructure. Imported natural gas is expected to supply both industry and power generation. Phase 1 of the project will use a 170,000-cu m floating storage unit attached to 3 million tpy of onshore regasification capacity. Phase 2 will add 220,000 cu m of onshore storage (potentially replacing the FSU) and 4.5 million tpy of regasification. ZET hopes to complete detailed engineering during 2027 to reach final investment decision in 2028 and start operations in 2030. Reuters reported last week that ExxonMobil Corp. had signed a preliminary deal to supply LNG to ZET. Developed as a joint between Vopak Terminal Durban and Transnet Pipelines, ZET project is expected to be South Africa’s first LNG terminal. The consortium will design, develop, construct, finance, operate, and maintain the terminal in the South Dunes Precinct at the Port of Richards Bay over a 25-year concession. EPC execution will be subject to ZET’s localization and economic development objectives. Successful contractors will be expected to support local supplier participation, skills development, and the use of local labor. Qualifying parties will be included in the project’s vendor database and may be shortlisted for subsequent phases as potential preferred contractors or subcontractors. The EOI submission window closes July 9, 2026. Interested contractors are invited to access the full EOI documentation here. South African utility Eskom and ZET earlier this month signed a head of agreement (HOA) establishing the framework for a long-term strategic partnership to support South Africa’s gas-to-power program, underpinning a planned 3-Gw power plant near the terminal in KwaZulu-Natal. Vopak Terminal Durban is owned by Royal Vopak and Reatile Group

Petrobras greenlights renewables plant for RPBC refinery

REDUC’s fist soybean oil-based SAF sale Announcement of FID on the RPBC renewables plant followed Petrobras’ June 17 confirmation that its 239,000-b/d Duque de Caxias (REDUC) refinery in the Baixada Fluminense area of Rio de Janeiro had completed first production and sale of a first 3,800-cu m batch of SAF made from soybean oil certified under the CORSIA low Land Use Change (ILUC) risk standard, which verifies sustainability criteria and a lower risk of impact on new land areas. Produced via co-processing and featuring 1% renewable content, the SAF batch marked “commercialization of the world’s first SAF made from certified low-ILUC-risk soy [to demonstrate] Petrobras’s commitment to sustainability, the energy transition, and the development of products aligned with market and societal demands [for lower-carbon solutions],” said Angélica Laureano, Petrobras’ director of logistics, sales, and markets. In October 2025, the REDUC refinery secured Brazil’s first international approval to advance commercial-scale production of SAF via the hydroprocessed esters and fatty acids (HEFA) co-processing route complying with ISCC System GmbH’s International Sustainability Carbon Certification (ISCC) standards, validating that SAF produced at the site meets the highest international sustainability and lifecycle carbon emission standards. Developed under ICAO’s CORSIA, the ISCC CORSIA certification was a prerequisite for commercial-scale SAF production following rigorous assessment of the production’s lifecycle carbon emissions and traceability. Equipped to produce as much as 10,000 b/d of SAF using a blend of conventional petroleum and up to 1.2% renewable feedstock, REDUC’s integration of bio-based oils—such as vegetable oil—into existing refining infrastructure via the HEFA co-processing method allows the refinery to produce SAF alongside conventional jet fuel with minimal investment, Petrobras previously said.

Equinor to expand Troll with TWIN subsea development

Equinor Energy AS and partners will invest about NOK 4 billion ($410 million) in the new Troll West increased gas recovery north (TWIN) subsea development in Troll field in the North Sea. The TWIN project consists of two wells in a template and a pipeline connected to existing subsea infrastructure. The umbilical and monoethylene glycol line will be extended to the new development. The project is expected to contribute about 11 billion std cu m of gas to Troll. It is the third step of Troll Phase 3, which produces gas from the Troll West reservoir. Recoverable reserves from Troll Phase 3, mainly gas, are estimated at 2.2 billion boe. In accordance with the Petroleum Act, the partnership will now send an announcement to the Ministry of Energy concerning the development. An environmental impact assessment has been carried out. Troll, which supplies as much as 10% of Europe’s daily demand for gas, contains about 40% of the total gas reserves on the Norwegian continental shelf and was developed in phases, with gas extraction from Troll Øst in Phase 1 and oil from Troll West in Phase 2. The oil in Troll West is produced from multiple subsea templates tied into Troll B and Troll C via pipelines. Production from the Troll C installation started in 1999. Troll C is also used for production from Fram, Fram H-Nord, and Byrding. Several amended development plans were approved in connection with installing multiple subsea templates on Troll West. Equinor Energy AS is operator of TWIN (30.55%) with partners Petoro AS (55.93%), A/S Norske Shell (8.19%), TotalEnergies EP Norge AS (3.69%), and ConocoPhillips Skandinavia AS (1.64%).

ICYMI: Upstream M&A slows on pricing gaps while deal appetite holds

Despite a slowdown in headline deal values this spring, upstream mergers and acquisitions remain active beneath the surface. In this ICYMI episode of the Oil & Gas Journal ReEnterprised podcast, Mikaila Adams, managing editor, examines data from Enverus and Rystad Energy detailing international and North American upstream deal markets in 2025 and into 2026. The discussion explores how pricing uncertainty widened the gap between buyers and sellers, creating a temporary pause rather than a collapse in market activity. The episode also looks at where capital continues to flow and what those trends reveal about the industry’s direction. From North American consolidation led by the Devon Energy–Coterra Energy merger to continued interest in gas-weighted assets tied to Gulf Coast LNG exports, the analysis highlights the forces shaping today’s upstream M&A landscape. It also considers the likelihood of additional divestitures, private equity activity, and asset sales as companies refine their portfolios, pointing to continued dealmaking momentum even in a more volatile market. References Devon, Coterra joining forces to create 1.6 million boe/d shale titan https://www.ogj.com/general-interest/companies/news/55354563/devon-coterra-joining-forces-to-create-16-million-boe-d-shale-titan Ovintiv to divest Anadarko assets for $3 billion https://www.ogj.com/general-interest/companies/news/55358241/ovintiv-to-divest-anadarko-assets-for-3-billion Insights: Vaca Muerta’s scale, productivity—and why it has more to give https://www.ogj.com/home/podcast/55370296/insights-vaca-muertas-scale-productivityand-why-it-has-more-to-give Mitsubishi to enter US shale gas business through Haynesville asset acquisition https://www.ogj.com/general-interest/companies/news/55344199/mitsubishi-to-enter-us-shale-gas-business-through-haynesville-shale-acquisition Shell to expand Canadian operations with $16.4-billion acquisition of ARC Resources https://www.ogj.com/general-interest/companies/news/55373597/shell-to-expand-canadian-operations-with-164-billion-acquisition-of-arc-resources US upstream M&A hits $38 billion in 1Q26 before volatility temporarily pauses the market https://www.enverus.com/newsroom/u-s-upstream-ma-hits-38-billion-in-1q26-before-volatility-temporarily-pauses-the-market/ International upstream M&A stuck at historic low https://www.enverus.com/newsroom/international-upstream-ma-stuck-at-historic-low/ Upstream deal value falls 83% as oil price uncertainty widens the buyer-seller gap https://www.rystadenergy.com/insights/upstream-deal-value-falls Iran war impact on global oil markets https://www.ogj.com/IranWar

JPMorgan conference notes: COO says EOG will ‘continue to be explorationist’

When Gaspar announced the $22 billion deal for Coterra in February, investors and analysts quickly began to question the future of the Marcellus assets that had been under Coterra’s umbrella. Activist investor Kimmeridge had been calling for Coterra’s board to divest that asset and focus on the Delaware, a push that has since landed on Gaspar’s desk and one the executive has repeatedly said will be addressed via a broader review of the enlarged Devon’s holdings. Several times during his chat with Jayaram, Gaspar touted Devon’s prowess in the Delaware—adding Coterra’s operations has grown its footprint there to nearly 750,000 acres—and delineated the review process as covering three main points. What’s the value of the various assets on their own? What’s the market for them and who might the strategic and financial buyers be? (Here, Gaspar specifically mentioned asset-backed securitization (ABS) money “that’s really entered the space.”) And thirdly, and “very fundamentally important,” how complementary are the individual business units to each other? Could discerning observers interpret the latter as suggesting that the Marcellus assets are indeed the odd duck in the group, as Kimmeridge has said? (See the map above.) And is the ABS reference more than a winking acknowledgment of a Reuters report a month ago that money manager Stone Ridge Asset Management had bid $8 billion for the Marcellus division using securitization as a big financial lever? Gaspar didn’t elaborate and Jayaram didn’t press the issue. But Gaspar emphasized that clarity around the review isn’t far away: “We’ve telegraphed this is more of a months exercise, not a years exercise. […] The view with which we are approaching this, we are aggressive. We will be mindful of how do we take this moment in time to create more value for the shareholders.”

You can’t build sovereign infrastructure with Broadcom, says CISPE

CISPE has cited several reasons why VCF doesn’t fit the bill, in particular highlighting its lack of portability. This means that it doesn’t qualify as resilient under CISPE’s Sovereign and Resilient Cloud Framework. Earlier this month, the EU unveiled proposals for its Cloud and AI Development Act (CADA) to strengthen Europe’s digital economy. CADA will encourage investment in European research, lay down conditions for European data centers, and provide a single EU-wide assessment framework for cloud and AI sovereignty. CISPE said that Broadcom is a long way short of fulfilling the conditions proposed for CADA. Broadcom would fail to meet anything but a Level 1 certification under the CADA sovereignty framework, CISPE said, adding that Broadcom’s terms and conditions offer limited maintenance commitments, no source-code escrow, no substitution plan and no Data Act certification, all likely to fall foul of CADA’s recommendations.

Break legacy lock-in: Strategic options for enterprises facing the vSphere 8 deadline

The acquisition of VMware by Broadcom has caused many enterprise IT leaders to reexamine their infrastructure strategies. For organizations running vSphere 8, the October 2027 end-of-support deadline is rapidly becoming a planning priority. What may appear to be a routine upgrade is driving bigger discussions about cost, flexibility, cloud strategy, and long-term infrastructure direction. Many organizations have not only begun evaluating alternatives but also are leaving VMware. “VMware has been a great, innovative company,” says Harsha Kotikela, senior director of product and solutions marketing at Nutanix. “But since the acquisition, their business model has fundamentally changed, and that is what is forcing IT leaders to adapt.” Sticker shock, vendor lock-in, and the need for flexibility One of the biggest catalysts has been licensing costs. Organizations that had grown accustomed to predictable contracts have encountered significant pricing increases, creating what Kotikela describes as “sticker shock.” At the same time, some enterprises are reevaluating their vendor relationships due to concerns about support availability and changes in partner engagement models. Beyond immediate operational concerns, IT leaders are also focused on future requirements. Hybrid cloud environments have become the norm, with applications and data distributed across data centers, public clouds, and edge locations. AI initiatives are adding another layer of complexity, requiring infrastructure that can support workloads wherever they need to run. “The future is about flexibility,” Kotikela says. “If enterprises want to implement AI at the edge, in the data center, or in the cloud, they need the capability to manage that environment without creating silos.” That flexibility is becoming a critical factor in infrastructure decisions. Organizations increasingly want platforms that support multiple deployment models, open APIs, and cloud-native technologies to minimize the risk of vendor lock-in. How a future-ready platform addresses IT and business requirements Nutanix positions its architecture around openness and choice, according to

Qualcomm’s $3.9 billion purchase of Modular aims to change the data center dynamic

“Nvidia has something like 85% of the AI accelerator chip market,” he pointed out. “Sure, they have nowhere to go but down, but that’s still going to take them a while. More importantly, they have literally spent decades working with practitioners in AI and ML and compute-intensive fields, indoctrinating them into their CUDA software ecosystem. Rewriting that tool chain will take institutional change at most organizations, which means years, if not decades, to uncouple.” “Organizations that think they’ve achieved agnosticism because they’re using high-level abstractions like PyTorch, well, they have come closest,” he observed. “But just cutting and pasting the same code into AMD Instinct can lead to memory and dependency errors. It’s like VM lift and shifts to the public cloud 10 years ago. Easier, but still possible to screw up.” Nonetheless, Annand said that the deal, if it goes through, is still good news for enterprises.

KKR Bets Big on AI Infrastructure With Helix Launch, Tapping Former AWS CEO Adam Selipsky to Build a New Hyperscale Model

To close industry watchers, it’s really no secret that the AI infrastructure race has entered another phase; one where capital formation itself may become as strategically important as GPUs, power procurement, or liquid cooling. And in launching Helix Digital Infrastructure, investment giant KKR is making a calculated wager that hyperscalers no longer simply need developers or financiers. They need a partner capable of orchestrating capital, energy, connectivity, and data center execution as a unified platform. The significance of that strategy is underscored by the executive chosen to lead it. Adam Selipsky, the former CEO of Amazon Web Services and one of the industry’s most experienced cloud operators, will serve as Co-Founder and CEO of Helix, bringing firsthand experience from the very class of customers the new venture intends to serve. A New Model for AI Infrastructure Helix launches with more than $10 billion in long-duration committed capital from founding investors including KKR, the Kuwait Investment Authority (KIA), NVIDIA, and Vistra. But the headline number tells only part of the story. The company has been structured around an increasingly important thesis: that AI infrastructure can no longer be assembled piecemeal. Rather than treating data centers, electrical supply, transmission capacity, and fiber connectivity as separate procurement exercises, Helix proposes a vertically coordinated approach in which a single organization manages and finances the entire infrastructure stack. According to KKR, the objective is to reduce execution risk and accelerate deployment for hyperscale customers facing unprecedented AI demand. As AI factories grow from hundreds of megawatts toward gigawatt-scale campuses, synchronization among land acquisition, utility planning, financing, construction, and technology deployment has emerged as one of the industry’s defining challenges. Helix is effectively positioning itself as an operating platform designed to simplify that complexity. Why Selipsky Matters The appointment of Adam Selipsky may be the announcement’s

Beyond Hyperscale: Why Enterprise Data Centers Still Matter in the AI Era

“The enterprise data centers, even the new ones, tend to be far, far smaller than new hyperscale deployments,” Killian said. “Not uncommon to see enterprises deploy a quarter meg or one meg or two, maybe up to 10 megs. Whereas the hyperscale guys are deploying 40 up to 300 meg facilities.” But scale alone does not tell the story. For every one of the roughly 20 hyperscale users that dominate headlines, Killian noted, there may be 50 to 100 times as many large and mid-sized enterprise users. Those companies run critical business systems, purchase hardware, software, telecom and services, employ large data center teams, and often operate multiple facilities across domestic, edge, EMEA and Asia-Pacific footprints. In other words, enterprise demand may be smaller in unit size, but it remains massive in aggregate. And as AI shifts from training to inference, the enterprise data center could become newly strategic. Enterprise AI Is Not Hyperscale AI Killian’s central point is that enterprise infrastructure requirements differ materially from hyperscale requirements. Hyperscalers are primarily optimizing for massive scale and speed to market. Enterprises, by contrast, tend to prioritize reliability, flexibility, integration into broader IT systems, and audit and compliance. That difference has major implications for developers and colocation providers. “The real industry opportunity is to take some of the innovation and the economies of scale that we’re seeing from the hyperscale builds to deliver smaller chunks of data center capacity,” Killian said. That might mean adapting lessons from 40 MW or 100 MW campuses into enterprise-ready deployments of 2 MW, 4 MW or 8 MW. Killian pointed to providers such as DataBank and Flexential as examples of companies working to deliver hyperscale-derived efficiencies in smaller enterprise increments. He also noted that QTS and other large campus developers may reserve portions of multi-building campuses

Revolutionizing Data Center Cooling: Innovations for AI and HPC Growth

This is a crucial point for AI infrastructure. In some markets, water can be as politically and operationally difficult as power. Evaporative cooling and cooling towers can consume large volumes of water, while discharge permits can slow projects or limit operations. Gradiant claims HyperSolved can expand access to alternative sources such as municipal reuse and impaired supplies, reduce reliance on freshwater, protect cooling performance through integrated treatment and AI-enabled operations, and minimize discharge through high-recovery concentration and reuse. The platform uses containerized systems for immediate or temporary capacity while also supporting permanent infrastructure and lifecycle operations from commissioning onward. That fits the AI data center buildout, where developers may need bridge capacity during construction, phased water infrastructure, or interim systems while permanent treatment plants are completed. This can address the speed of deployment issue that plagues many data center solutions. Water is becoming a siting and scaling variable that has to be addressed. A site may have land and power prospects, but if water sourcing, reuse, or discharge cannot be solved, the project will face higher costs, delays, and local opposition. Gradiant is positioning itself as the managed water layer for hyperscale AI, similar to how power providers, cooling vendors, and network suppliers each own critical infrastructure domains. The Pattern: Hybridization, Standardization, and Industrial Scale The announcements included here make it clear that cooling is seeing significant attention from technology vendors, and not just state-of-the-art new technologies such as direct-to-chip, but also traditional data center air cooling. T-Global and SiPearl are working on high-conductivity materials and two-phase modules for HPC chips. Castrol is providing fluids for direct-to-chip and immersion environments. These are technologies aimed at the heat source itself, where higher chip power and rack density are overwhelming conventional approaches. The reference design offerings from Johnson Controls acknowledges the importance

Microsoft will invest $80B in AI data centers in fiscal 2025

And Microsoft isn’t the only one that is ramping up its investments into AI-enabled data centers. Rival cloud service providers are all investing in either upgrading or opening new data centers to capture a larger chunk of business from developers and users of large language models (LLMs). In a report published in October 2024, Bloomberg Intelligence estimated that demand for generative AI would push Microsoft, AWS, Google, Oracle, Meta, and Apple would between them devote $200 billion to capex in 2025, up from $110 billion in 2023. Microsoft is one of the biggest spenders, followed closely by Google and AWS, Bloomberg Intelligence said. Its estimate of Microsoft’s capital spending on AI, at $62.4 billion for calendar 2025, is lower than Smith’s claim that the company will invest $80 billion in the fiscal year to June 30, 2025. Both figures, though, are way higher than Microsoft’s 2020 capital expenditure of “just” $17.6 billion. The majority of the increased spending is tied to cloud services and the expansion of AI infrastructure needed to provide compute capacity for OpenAI workloads. Separately, last October Amazon CEO Andy Jassy said his company planned total capex spend of $75 billion in 2024 and even more in 2025, with much of it going to AWS, its cloud computing division.

John Deere unveils more autonomous farm machines to address skill labor shortage

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Self-driving tractors might be the path to self-driving cars. John Deere has revealed a new line of autonomous machines and tech across agriculture, construction and commercial landscaping. The Moline, Illinois-based John Deere has been in business for 187 years, yet it’s been a regular as a non-tech company showing off technology at the big tech trade show in Las Vegas and is back at CES 2025 with more autonomous tractors and other vehicles. This is not something we usually cover, but John Deere has a lot of data that is interesting in the big picture of tech. The message from the company is that there aren’t enough skilled farm laborers to do the work that its customers need. It’s been a challenge for most of the last two decades, said Jahmy Hindman, CTO at John Deere, in a briefing. Much of the tech will come this fall and after that. He noted that the average farmer in the U.S. is over 58 and works 12 to 18 hours a day to grow food for us. And he said the American Farm Bureau Federation estimates there are roughly 2.4 million farm jobs that need to be filled annually; and the agricultural work force continues to shrink. (This is my hint to the anti-immigration crowd). John Deere’s autonomous 9RX Tractor. Farmers can oversee it using an app. While each of these industries experiences their own set of challenges, a commonality across all is skilled labor availability. In construction, about 80% percent of contractors struggle to find skilled labor. And in commercial landscaping, 86% of landscaping business owners can’t find labor to fill open positions, he said. “They have to figure out how to do

2025 playbook for enterprise AI success, from agents to evals

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More 2025 is poised to be a pivotal year for enterprise AI. The past year has seen rapid innovation, and this year will see the same. This has made it more critical than ever to revisit your AI strategy to stay competitive and create value for your customers. From scaling AI agents to optimizing costs, here are the five critical areas enterprises should prioritize for their AI strategy this year. 1. Agents: the next generation of automation AI agents are no longer theoretical. In 2025, they’re indispensable tools for enterprises looking to streamline operations and enhance customer interactions. Unlike traditional software, agents powered by large language models (LLMs) can make nuanced decisions, navigate complex multi-step tasks, and integrate seamlessly with tools and APIs. At the start of 2024, agents were not ready for prime time, making frustrating mistakes like hallucinating URLs. They started getting better as frontier large language models themselves improved. “Let me put it this way,” said Sam Witteveen, cofounder of Red Dragon, a company that develops agents for companies, and that recently reviewed the 48 agents it built last year. “Interestingly, the ones that we built at the start of the year, a lot of those worked way better at the end of the year just because the models got better.” Witteveen shared this in the video podcast we filmed to discuss these five big trends in detail. Models are getting better and hallucinating less, and they’re also being trained to do agentic tasks. Another feature that the model providers are researching is a way to use the LLM as a judge, and as models get cheaper (something we’ll cover below), companies can use three or more models to

OpenAI’s red teaming innovations define new essentials for security leaders in the AI era

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More OpenAI has taken a more aggressive approach to red teaming than its AI competitors, demonstrating its security teams’ advanced capabilities in two areas: multi-step reinforcement and external red teaming. OpenAI recently released two papers that set a new competitive standard for improving the quality, reliability and safety of AI models in these two techniques and more. The first paper, “OpenAI’s Approach to External Red Teaming for AI Models and Systems,” reports that specialized teams outside the company have proven effective in uncovering vulnerabilities that might otherwise have made it into a released model because in-house testing techniques may have missed them. In the second paper, “Diverse and Effective Red Teaming with Auto-Generated Rewards and Multi-Step Reinforcement Learning,” OpenAI introduces an automated framework that relies on iterative reinforcement learning to generate a broad spectrum of novel, wide-ranging attacks. Going all-in on red teaming pays practical, competitive dividends It’s encouraging to see competitive intensity in red teaming growing among AI companies. When Anthropic released its AI red team guidelines in June of last year, it joined AI providers including Google, Microsoft, Nvidia, OpenAI, and even the U.S.’s National Institute of Standards and Technology (NIST), which all had released red teaming frameworks. Investing heavily in red teaming yields tangible benefits for security leaders in any organization. OpenAI’s paper on external red teaming provides a detailed analysis of how the company strives to create specialized external teams that include cybersecurity and subject matter experts. The goal is to see if knowledgeable external teams can defeat models’ security perimeters and find gaps in their security, biases and controls that prompt-based testing couldn’t find. What makes OpenAI’s recent papers noteworthy is how well they define using human-in-the-middle

Stay Ahead, Stay ONMINE