Can’t a multi-trillion-dollar company have secrets any more?

Dozens of cryptographically verified open source packages from Microsoft were compromised late last week to add advanced credential-stealing code that was triggered when developers opened them in AI coding agents.

In all, multiple researchers said, 73 packages were flagged as malicious when automated systems on GitHub blocked them on the platform. Rather than noting they are malicious—and that developers who used AI agents to work with them should assume their systems are compromised—the Microsoft-owned GitHub said it disabled the packages “due to a violation of GitHub’s terms of service.” The text went on to encourage the package owner to contact GitHub.

Devs: Assume compromise and proceed accordingly

It wasn’t until Monday that Microsoft even raised the possibility the packages were infected. In an email, the company stated: “We have temporarily removed some repositories as we investigate potential malicious content.”

The incident is the second supply-chain attack in as many months to breach an official Microsoft repository account. In mid May, the firm StepSecurity documented the compromise of Microsoft’s durabletask Python SDK on PyPI. The package is a framework for building fault-tolerant workflows and orchestrations to automate distributed transactions and other workflows. It receives 400,000 downloads per month.

The compromise packages executed a 28 KB payload that steals credentials from AWS, Azure, GCP, Kubernetes, password managers, and over 90 developer tool configurations. It then spreads laterally through cloud infrastructures to infect other developer machines. The attack, which has been linked to a threat actor tracked as TeamPCP, poisoned the durabletask package after compromising Microsoft credentials for publishing the package. The technique allows attackers to bypass the repository’s build pipeline entirely.

The malware used in the attack is tracked as Miasma. It’s essentially a clone of TeamPCP’s Mini Shai-Hulud toolkit, which the threat actor open-sourced recently. Security firm Cloudsmith said the malware harvests OIDC (OpenID-Connect) token credentials that are used in SLSA (Supply-chain Levels for Software Artifacts) provenance attestation, a method for providing cryptographically signed guarantees of a software’s integrity.

As was the case in the May compromise of Microsoft’s durabletask, the one last week made use of the functionality to steal a legitimate Microsoft OIDC token. It was also used in a separate supply-chain attack poisoning dozens of Red Hat packages.

New York City was the backdrop of this year’s IEEE Honors Ceremony, held on 24 April.

The event celebrates engineering pioneers who have developed technologies that have changed how people connect and learn about the world. This year’s celebrants included the engineers behind innovations such as text-to-donate technology, AI-powered diagnostic tools, and the graphics processing unit, among many others.

Prior to the Honors Ceremony, IEEE hosted a forum on 23 April for a select group of early-career achievers to exchange ideas and experiences with laureates and awardees, speakers, and IEEE leaders. Attendees from around the world, working in a variety of technical areas, shared their journeys and explored the intersections of technologies, disciplines, and missions.

The event culminated in Friday evening’s black tie Honors Ceremony, where IEEE celebrated medal laureates, including Jensen Huang, who received IEEE’s highest recognition, the IEEE Medal of Honor. Huang is a cofounder of Nvidia and its chief executive.

“IEEE has always been a home to those who see the future before others see it,” Mary Ellen Randall, IEEE president and CEO, said in her welcome speech.

Video highlights and photos from the event are available on the IEEE Awards website.

Exploring mission-driven tech and AI in art

Friday morning began with a conversation between Randall and Marian Croak, the recipient of this year’s IEEE Founders Medal. Croak was honored for “leadership in communication networks, including acceleration of digital equity, responsible artificial intelligence, and the promotion of diversity and inclusion.”

Croak, who serves as vice president of engineering at Google, headquartered in Mountain View, Calif., pioneered Voice over Internet Protocol (VoIP) technologies. When a person speaks into a telephone, VoIP converts their voice into digital signals that are transmitted over the Internet rather than traditional phone lines. Her work enabled audio and video conferencing. She also developed text-to-donate technology to raise money for those affected by Hurricane Katrina, which devastated New Orleans in 2005. The technology enables customers to donate money to a charity via their mobile service provider, which then bills them.

“Empathy has always been a driving force in the engineering that I’ve done,” she said.

She shared advice on how to stay creative: “Get out of the office. Go to an art museum, exercise, or play with children.” Croak said her grandchildren inspire her.

The daytime program also spotlighted AI’s use in the visual arts. Kathleen Kramer, the 2025 IEEE president, interviewed artist Refik Anadol, who is scheduled to open an AI art museum on 20 June in Los Angeles. Dataland’s exhibits are powered by an open-access model developed by Anadol’s studio.

For the museum’s first exhibition, “Machine Dreams: Rainforest,” the model collected visual data about the natural world from the Smithsonian National Museum of Natural History, London’s Natural History Museum, and the Cornell Lab of Ornithology, with their permission. The information, including up to a half billion images, will form the basis for a variety of AI-produced art, Anadol said.

Anadol said he was inspired to mix AI with art by the movie Blade Runner. He said he believes “machines can become collaborators,” as “data is a form of pigment.”

Data also plays an important role in the work of artist and author Giorgia Lupi. The artist is a partner at design firm Pentagram.

Lupi said she uses data to tell stories, including chronicling her struggles with a chronic illness.

“Data is an abstraction of our reality,” she said.

One of her recent projects, “A Data Love Letter to the Subway,” was shown last year in the Dey Street Passageway in New York City. The video was made using data from the Metropolitan Transportation Authority about each train line, including timetables, ridership, and people’s travel habits. Based on the information Lupi gathered, she documented how commuters traveling on different subway lines encountered one another without realizing it.

By exploring data on this year’s IEEE award recipients, she collaborated with IEEE to create an animated video illustrating the shared pathways and collaborations among the honorees. It debuted at the Honors Ceremony.

Honoring engineering giants

The Honors Ceremony, held at Cipriani 42nd Street, recognized more than 20 laureates and innovators.

More than 92 million selfies are taken worldwide every day, PhotoAiD estimates. A selfie wouldn’t be possible without Eric Fossum’s invention of the CMOS image sensor. Developed at NASA’s Jet Propulsion Laboratory, in Pasadena, Calif., the “camera on a chip” was intended for use in space, but it is now found in smartphones, medical devices, and vehicles. Fossum, an IEEE Life Fellow, received the IEEE Jun-ichi Nishizawa Medal, which recognizes outstanding contributions to materials and device science and technology.

“Engineering is a pursuit of what must be possible. [IEEE is] the spirit, the conscience, of our profession.” —Jensen Huang, founder and CEO of Nvidia

The medal, he said, “is at the top of the IEEE staircase of being recognized by your peers.”

The IEEE Holonyak Medal for Semiconductor Optoelectronic Technologies went to Steven P. DenBaars, a professor of materials and electrical and computer engineering at the University of California, Santa Barbara. DenBaars was honored for his work in semiconductors, which laid the foundation for high-resolution LED and laser displays, modern solid-state lighting, and more.

“This work has always been a team effort…I’m excited and curious about the role gallium nitride micro LEDs will play in optical communications,” he said in his acceptance speech.

The ceremony ended with the Medal of Honor presentation to Huang, who received a standing ovation. He was recognized for his “leadership in the development of graphics processing units and their application to scientific computing and artificial intelligence.”

The IEEE honorary member donated his cash prize to IEEE TryEngineering, which provides teachers with a library of lesson plans and offers educational summer camps. The Jen-Hsun and Lori Huang Foundation matched his gift, and the additional donation is destined to fund scholarships for new graduates.

“Engineering is a pursuit of what must be possible. [IEEE is] the spirit, the conscience, of our profession,” Huang said.

In recent days, founders and founders-turned-investors took to X to share horror stories about being mistreated by VCs. Their complaints ranged from VCs falling asleep during pitch meetings to investors suggesting a founder fire a co-founder.

Brendan Foody, co-founder of the AI talent platform Mercor, which was last valued at $10 billion, went so far as to call out Sequoia, arguably one of the most elite VC firms in the world.

“The “sequoia scam” is worse than a single horror story,” Foody wrote on X. “in the last 6 [months] ive seen a half dozen rounds where sequoia invests in 2 tranches. everyone pretends they only did the higher valuation. founders misrepresent this to their employees & then shop it to angels too.”

TechCrunch has previously reported on VCs investing in the same round at different valuations. Under this mechanism, the lead VC firm invests a significant chunk of its capital at a lower, preferential valuation, while putting a much smaller portion of capital in at a drastically higher price. The massive “headline” valuation that gets announced manufactures the perception of a dominant market winner, masking the fact that the lead investor’s actual average entry price was significantly lower.

The disparity can be stark. For example, when the AI-driven IT helpdesk startup Serval announced a $75 million Series B at a $1 billion valuation, the announcement didn’t tell the whole story. According to The Wall Street Journal, Sequoia’s actual lowest entry point valued the company at just $400 million — less than half the headline figure. The gap between those two numbers is the gap between perception and reality that Foody is pointing at.

Serval isn’t alone. At Aaru, a startup that uses AI to simulate user behavior for market research, lead investor Redpoint backed the company at a $450 million valuation despite an announced $1 billion headline price.

Sequoia’s Shaun Maguire pushed back on Foody’s characterization directly. “TBH I have seen some of this behavior but I think it’s unfair to call it the ‘Sequoia scam,’” Maguire wrote in response to Foody on X. “This has happened approximately five times during my seven years at Sequoia. What happens is other investors are willing to pay a high price for a hot company — usually AI — at multiples above what we’re willing to pay. So we try to decouple the company-building relationship with our partner from the capital, and this leads to two tranches at different valuations in close succession.

“I’m not aware of anything shady here,” Maguire continued, “but if you’ve seen it I’d love to know. VC is a repeated game, so it just doesn’t make sense for us to try to mislead people. And if anyone has, I’d love to know. And in general, congrats on the success of Mercor — it was a miss for us.”

Maguire’s response frames the practice as a market reality rather than a deliberate maneuver — Sequoia, he suggests, is simply unwilling to pay what competitors will pay for the hottest deals, so it structures its participation differently. Whether that explanation fully holds up depends on a question Maguire doesn’t address: what founders are telling the people who don’t already know about the lower tranche.

Although Sequoia appears to use this pricing mechanism most frequently, Foody acknowledged it isn’t the only firm using this tactic. And while the dual-pricing structures certainly inflate a startup’s perceived worth and help attract top talent, calling the practice a “scam” may be going too far.

That’s because the stock options granted to employees should theoretically be valued based on the blended price of all tranches, said Jason Woo, partner in valuation and financial modeling at Armanino. Woo’s group provides the independent 409A valuations that startups use to price employee stock options. A 409A is an independent appraisal that’s supposed to set the fair market value of a private company’s shares — it’s the number that determines what employees actually pay for their options, regardless of what valuation gets announced in a press release.

There’s a catch, however: 409A valuations are widely understood to skew low. Because the 409A sets the strike price of employee options — and a lower strike price means a smaller tax bill for the company — there is a structural incentive to keep that number down. Which means the independent appraisal that’s supposed to protect employees from an inflated headline valuation is also, by design, not trying particularly hard to reach the top of the range. Employees may not be paying the headline price for their options, but they may not be getting the full picture either.

The angel question is more complicated. Unlike employees, angels are writing checks, not receiving options. There is no independent appraiser standing between an angel investor and whatever number a founder chooses to share.

The dual-pricing structure is just one of way VCs and founders game the perception of success in a hyper-competitive market. Another, more pervasive tactic involves manipulating or outright overstating annual recurring revenue (ARR).

The VC Niko Bonatsos, a longtime veteran of General Catalyst who more recently founded Verdict Capital, addressed this issue during one of TechCrunch’s events in Athens last month. “We [at Verdict] mostly invest before metrics, before product, before the company [has fully taken shape] but I do have a past portfolio, and sometimes the conversations are telling. I’ll get a call or an email with a very high ARR number. I’ll think: I didn’t remember that company doing so well. So I reach out to the founder: ‘What happened? Why are the numbers so strong?’ And the answer is: ‘Oh yeah, it’s 365 times the revenue we made yesterday because one of our campaigns hit.’ So yeah, some of these terms have lost meaning.”

Foody declined to comment further. Sequoia didn’t immediately respond to a request for comment.

— With additional reporting from Connie Loizos

President Claudia Sheinbaum drove immediately into the stage set up inside a Mexican air force hangar in Mexico City, giving the world its first glimpse of the Olinia Uno, a 100% Made-in-Mexico electric vehicle. The project was spearheaded by a team of Mexican engineers and academics who worked tirelessly to create a vehicle that could help propel the country into the electric-vehicle era. The Olinia Uno is a compact six-seater van that starts at 150,000 pesos, or around $8,000 to $8,600 USD at current exchange rates. Its intended demographic is, as expected, people who take short trips about the city, which is exactly what most driving in Mexico’s cities involves.

The vehicle adopts a no-nonsense design, pairing a boxy shape with a high roof to maximize interior space and ease of access.. There are multiple wide windows along the sides and back, which improve the driver’s visibility and allow the passengers to enjoy some natural light inside. The styling is quite apparent, with a solid basic two-tone white and black scheme that looks clean and utilitarian, and because the doors open wide and have grab handles on the side, getting in and out of the van is a snap for families or anyone with a lot of gear to haul around.

Sale

Hiboy S2 Electric Scooter – 8.5″ Solid Tires – Up to 17 Miles Long-Range & 19 MPH Portable Folding…

• Powerful Motor And Long Battery Life – The 350W electric brushless hub motor allows for speeds up to 19mph. High capacity battery(36V 7.8AH) with a…
• Stable And Durable For Commuting – Extremely smooth when accelerating and turning, Hiboy electric scooter have been made of solid and durable quality…
• Being Safer With Technology – Hiboy Electric Scooter Adults S2 is made with front, middle, and rear 3 lights. The double braking system features a…

Engineers picked a 13-kilowatt electric motor and a 14.7-kilowatt-hour lithium iron phosphate battery to power the van, which should have a top speed of roughly 50 kilometers per hour and a range of more than 100 kilometers on a standard city trip. The Olinia Uno also offers standard safety features like front disc brakes and electronic power steering to make it stable at low speeds, as well as a reverse camera and LED lamps for enhanced safety.

Charging the Olinia Uno is simple; users simply connect it into an ordinary household outlet like any other appliance. The recharge time will be approximately 4 hours if you have 220-volt energy and up to 8 hours if you have 110-volt power. This requires neither a pricey wall box nor a public fast charger. And the running expenses are modest. After accounting for gasoline, maintenance, and reliability, it costs around 0.50 pesos per kilometer, which is much less than a gas-guzzling cab or even many motorcycles.

The interior is meant to seat six people, all with enough seatbelts, and because the designers took wheelchair users into account, there is enough room to accommodate a full-sized wheelchair without having to fold it up. There is also adequate inside lighting to help with nighttime boarding. The dashboard has a 7-inch central screen that shows speed, some basic gauges, and media controls. Bluetooth 5.0 connectivity, USB and USB-C ports, power windows, and central locking are all standard. The canopy totally protects against rain and heat, giving it a substantial benefit over an open motorcycle for day-to-day transportation as safely as possible.

It took 18 months to bring the Olinia Uno to market, with the help of many public colleges, research organizations, and over 80 Mexican academics and engineers. The plan is for 50% of the parts to be made in Mexico from the start, with the goal of increasing to 75% by 2030. They’re opening an assembly plant in Puebla later this year, with plans to increase production to 20,000 units per year by 2027. If that wasn’t enough, they’re also planning to install 2,000 public charging points in Mexico City, the State of Mexico, and Puebla to make it even easier for people to get behind the wheel of the Olinia Uno or set up taxi fleets.
[

Apple confirmed at WWDC that its Foundation Models aren’t a cut and paste of Gemini, and are all-Apple through-and-through. We’ve been telling you this all along.

WWDC 2026 was a rollercoaster event that primarily focused on Apple’s latest AI upgrades. This time, though, they had a little help from Google.

Apple pre-announced that Google was providing Gemini technology to help develop the new Apple Foundation Models, but didn’t say much else. The rampant speculation painted a portrait of scrambling and failure, as usual.

The reality is exactly what AppleInsider has been reporting all along, which makes sense. All of the information was there if you were willing to see it for what it was.

Simply, it’s this:

The upgraded Apple Foundation Models power a new Siri AI and Apple Intelligence that utilize private, safe, and secure on-device and Private Cloud Compute server-side operation. The new models were built with the aid of Google Gemini and its technologies, but through distillation and training, not full replacement.

Apple has confirmed the end result is pure Apple technology and code. When you interact with Apple Foundation Models, you never touch a drop of Google code, Gemini agents, or even Google Search.

It’s Apple software all the way down.

Apple’s new models

A short talk was held with a few Apple executives after the main WWDC 2026 keynote. Apple SVP of Software Engineering, Craig Federighi, was joined by Sebastien Marineau-Mes, Mike Rockwell, and Amar Subramanya on stage to discuss the new Apple Foundation Models (AFM).

Apple’s executive gaggle at WWDC

Here’s what was shared:

The on-device models are AFM Core and AFM Core Advanced. The advanced version is natively multi-modal with a sparse architecture that enables more capable features without leaving your device.

AFM Cloud is the Private Cloud Compute base model that handles more taxing AI requests that can’t be run on-device. The AFM Cloud Image model is for image generation and editing.

Each of these four models is custom-built for Apple Silicon, trained using proprietary data, and further polished with distillation from the loaned Gemini models.

Then there’s AFM Cloud Pro that will be used for agentic tools and the most demanding tasks. It is going to use infrastructure provided by Google’s cloud servers and NVIDIA’s GPUs while remaining Private Cloud Compute certified.

Third parties can review the servers used by Apple to independently verify Private Cloud Compute certification. Basically, this means external entities can prove whether Apple is keeping or mishandling user data in its AI servers.

What the rumors got wrong

Apple did flub its initial Apple Intelligence rollout. It overpromised on features that would never be able to perform as expected due to AI hallucination rates and Siri’s ML-based backend.

iPhone 17 Pro Max will benefit from Apple’s most powerful models

That slow launch and delay in early 2025 led people to believe Apple would never be able to deliver “good AI.” Meanwhile, the world saw the grift grow as other companies promised world-changing or world-ending features that amounted to slop generators and nudify apps.

As AI sentiment waned, Apple powered through and had one record-breaking quarter after another without a strong AI offering. Even so, pundits claimed Apple’s next AI announcement would inevitably be a whiff unless they gave up and relied on someone else’s models.

They seemingly got their wish when Apple shared that it was partnering with Google to utilize Gemini technologies as the foundation of Apple Foundation Models. The pundits saw what they wanted to see and ran with it — Apple had given up and Apple Intelligence would now be Gemini.

Whether they called it “white label Gemini” or suggested Gemini agents would operate alongside Apple’s weaker models, they universally couldn’t imagine a world where Apple simply didn’t use Google in its AI revamp. Obviously, they were wrong.

Apple and Google may have been deliberately opaque in their press release around this partnership. One thing was clear, though.

Apple clearly said at the time that Siri and Apple Intelligence would be powered by Apple Foundation Models.

And on Monday, June 8, they exposed exactly that.

Many adults who spent hours navigating Mario through Dinosaur Land on their old Super Nintendo still like seeing the red-capped plumber ride alongside Yoshi. LEGO has transformed this throwback to the past into an actual model, called LEGO Super Mario World: Mario & Yoshi (set 71438), priced at $104 (was $130), that can be built, displayed, and even interacted with.

The finished model is around fifteen and a half inches tall and ten inches wide, perfectly replicating the blocky appearance of the original 16-bit sprites. Mario is perched high on Yoshi’s back, where he prefers to be, with his cape flowing behind him. Yoshi has been designed to match, with a green body and a white tummy / egg pouch that are perfectly blocky and sharp, just like in the original game.

Sale

LEGO Super Mario World: Mario & Yoshi Building Set for Adults, Ages 18+ – Gift Idea for Birthdays – DIY…

• LEGO Super Mario collectible set for adults – Return to a nostalgic era of side-scrolling Nintendo games with this brick-built Super Mario World…
• Pixelated Nintendo figures – Turn the handle at the base of the model to make Yoshi run with Mario on his back, and turn the dial behind Yoshi’s…
• Interactive experience – The set includes an Action Tag so you can also add LEGO Mario, LEGO Luigi or LEGO Peach (figures not included) and see…

The entire thing can be moved by turning a handle on the base’s side. Yoshi’s legs move in a running stride, while Mario bounces up and down in time with each step. It’s really smooth and enjoyable to see, comparable to what the duo used to accomplish on your screen years before. You may control Yoshi’s tongue by rotating a little dial on his head. Give it a turn, and the red tongue emerges before snatching it back with another turn. It’s not much, but it’s nice to witness one of Yoshi’s most well-known moves in action.

The build takes 1215 pieces, which are neatly divided into 15 numbered bags and include a full instruction booklet. Builders say this one is a lot of fun but not too challenging; the game’s pixelated style appears to flow naturally as you add layer after layer of plates and tiles. If you prefer to follow along on a screen, the official LEGO app has a digital version of the instructions.

The box includes an Action Tag, which is essentially an extra layer for those who already own some of the other Super Mario LEGO figures. Place one of the tiny Mario, Luigi, or Peach figures near your build and scan it with the app’s Action Tag to trigger a range of hilarious digital reactions. It’s a really unique way to connect the main display piece to your smaller interactive figures.

Contact information, direct messages and connected accounts were all potentially compromised, Meta said.

Hackers used Meta AI to hack into 20,225 Instagram accounts, Meta reported in a US local government data breach notice on 5 June.

According to the notice to the attorney general for Maine, the breach occurred on 17 April, but wasn’t discovered by the company until more than a month later, on 31 May.

The company explained that hackers exploited a now-resolved bug in its AI-assisted support tool designed to help Instagram users access their account after being locked out.

“HTS (high touch support) is an AI-assisted support tool designed to help users who are locked out of their Instagram accounts regain access,” said Amber Hannah, Meta’s associate general counsel for incident response.

“Users can request support from HTS and, as part of that process, can ask that a password reset link be sent to their email address.

“The tool itself worked properly and functioned as intended; however, due to a bug in a separate code path, the system did not properly verify that the email address provided by the individual requesting a password reset matched the email address associated with that user’s Instagram account.”

The bug allowed hackers to avoid triggering Instagram’s automated account protections, enabling password reset links to be sent to an email not connected to the account. Bad actors were then able to reset passwords to gain access to victims’ accounts if they did not have two-factor authentication enabled.

The hack affected prominent figures’ accounts, including the inactive Instagram handle for the Obama-era White House, beauty retailer Sephora and a senior US Space Force official.

Meta said that hackers could have potentially accessed sensitive data, including contact information, direct messages and communications, and connected accounts and linked services, such as email IDs. The company said that it would fix the bug before relaunching the AI tool.

In 2024, the Irish Data Protection Commission (DPC) fined Meta €251m for a 2018 data breach affecting approximately 29m Facebook accounts. The same year, the watchdog fined Meta €91m for improperly storing passwords.

In 2023, the company was fined €1.2bn by the DPC for violating GDPR guidelines by transferring users’ personal data outside of the EU.

AI-enabled cybercrime is fast becoming a sore point for companies, as attacks become more frequent and sophisticated. Just last month, hackers stole 8TB of data from the Taiwanese electronics manufacturer Foxconn, while medical equipment manufacturing giant Stryker was hit by a global cyberattack in March.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

A joint research collaboration between researchers at the University of Illinois at Urbana-Champaign (UIUC), UC Berkeley, and the open source AI-native vector database platform Chroma unveiled Harness-1, a 20-billion parameter open-source search agent built atop OpenAI’s gpt-oss-20B open source model that fundamentally redesigns how AI executes complex retrieval tasks.

Harness-1 achieves a massive leap in performance, scoring 73% average on its ability to recall relevant information correctly from a curated dataset, outperforming even GPT-5.4 (70.9%) and the next, most accurate open source search agent, Tongyi DeepResearch 30B, by 11.4 percentage points. (While GPT-5.5 has also been out for more than a month, the researchers didn’t test against this model as it wasn’t available when they were building theirs.)

Crucially for developers, the model and its environment are available immediately under the highly permissive Apache 2.0 license and model code/weights on Hugging Face.

Harness-1 also serves as proof-of-efficacy of another effort, Tinker, the distributed, web-based AI model training and fine-tuning API developed by Thinking Machines. Tinker was used specifically to train and run inference for Harness-1, highlighting how interactive infrastructure is actively enabling the next generation of autonomous models.

So how did the researchers do it?

Benchmarks Decoded (and Why Harness-1 Could Help Enterprises Tremendously)

To actually put these models to the test, the researchers evaluated Harness-1 and its competitors across eight highly complex search benchmarks. Rather than asking simple trivia questions, these tests required the AI to act like a real researcher sifting through diverse, dense data sources.

The benchmarks spanned several different domains, including open web searches, complex financial filings from the SEC, technical patent databases from the USPTO, and “multi-hop” question-answering tasks where the AI had to logically piece together scattered clues from multiple different documents to arrive at the correct answer.

When the results came in, Harness-1 dominated the open-source competition in its ability to successfully find and curate the right facts. Even more impressively, this relatively small 20-billion parameter model went toe-to-toe with massive, expensive proprietary AI systems. It actually outperformed heavyweights like GPT-5.4, Sonnet-4.6, and Kimi-K2.5 — thought to be the hundreds of billions or trillions of parameters. Only one giant frontier model—Opus-4.6 — managed to narrowly edge it out in overall average performance.

Harness-1 achieves its performance gains by offloading the exhaustive “bookkeeping” of a search session out of the model’s working memory and into a structured software environment.

As enterprise use cases grow more sophisticated, demanding that models autonomously sift through thousands of corporate documents or financial filings, these systems frequently succumb to “search amnesia”—forgetting their original queries, looping over rejected documents, or losing track of the specific claims they are trying to verify.

Until now, the prevailing solution to this amnesia has been brute force. Engineers typically force models to constantly reread an ever-expanding, append-only transcript of their own actions, piling every search, read, and thought back into a massive context window.

Harness-1 introduces a paradigm shift away from this method, proving that the bottleneck for true artificial autonomy isn’t necessarily the size of the model, but how efficiently its working environment manages state. It highlights once more, as Anthropic’s Claude Code has also done, that the raw model is arguably less important than the harness — or set of conditions — through which it runs.

Technology: Doing the Paperwork in the Environment

To understand the technical leap of Harness-1, consider a real-world analogy.

Imagine hiring a brilliant research assistant and placing them in an empty room without a desk, notepads, or filing cabinets. You ask them to write a comprehensive report on a highly complex topic, which requires them to read dozens of books while keeping every single quote, citation, and dead-end search perfectly memorized in their own head. Eventually, no matter how intelligent the assistant is, their cognitive load will max out, and they will start dropping facts or losing the thread of the assignment.

This is exactly how traditional search agents operate today. They are trained as policies over growing transcripts, meaning the model searches, reads, searches again, and appends everything into its own context window.

As lead researcher Patrick (Pengcheng) Jiang of the University of Illinois noted on X: “At some point the model is not just ‘searching’ anymore. It is also being asked to be a memory system, a note taker, a verifier, and a librarian.”

Harness-1 solves this by giving the AI a desk and a filing cabinet—what the research team calls a “state-externalizing harness.”

This harness is an active, surrounding environment that takes over the routine bookkeeping, maintaining a recoverable working memory that includes a candidate pool of documents, an importance-tagged curated evidence set, compact evidence links, and verification records.

By separating semantic choices from structural state management, the AI is freed up to do what it does best.

The policy still decides what to search, determines which documents to keep, and knows when to stop, while the environment simply holds the state.

Here is a subsection breaking down the training methodology and how it differs from prior agentic search models:

Training Harness-1: A Masterclass in Data Efficiency

The training pipeline for Harness-1 represents a fundamental shift in how the AI industry approaches agentic learning.

Historically, developers have treated search agents as policies operating over massive, ever-growing transcripts, forcing reinforcement learning (RL) algorithms to simultaneously optimize both semantic reasoning and the raw memorization of a search state.

Harness-1’s creators took a radically different approach: because their custom “harness” handles all the routine bookkeeping—like maintaining evidence links, candidate pools, and verification records—the training process only needed to teach the model how to operate this structured interface.

This division of labor drastically simplified what the underlying 20-billion parameter model actually needed to learn.

The process began with a remarkably narrow Supervised Fine-Tuning (SFT) stage. Rather than scraping petabytes of new behavioral data, the team generated just 899 filtered trajectories using a GPT-5.4 teacher agent that was plugged into the exact same harness environment the student model would eventually use.

The goal of this SFT phase was not to inject vast amounts of domain knowledge into the model, but simply to teach it the mechanical rhythms of a good researcher: how to format tool calls, how to tag documents by importance, and the discipline of verifying a claim before promoting it to the final curated set.

Following SFT, the model underwent Reinforcement Learning (RL) using an algorithm called CISPO, applied over full search episodes capping at 40 turns.

The team designed a highly specific terminal reward function that explicitly separated discovery from selection. The model was rewarded not just for finding a relevant document, but for successfully promoting it into the final answer set, while being penalized if it found the answer but failed to curate it.

The researchers also instituted a “tool diversity” bonus; without this specific incentive, they found the policy would quickly collapse into a lazy, search-heavy strategy where it spammed queries but bypassed the harder work of reading and verifying the text.

What makes Harness-1 truly innovative compared to prior work is its unprecedented data efficiency. The entire model was trained on roughly 4,400 unique items—899 SFT trajectories and 3,453 RL queries.

In stark contrast, competing open-source models required vastly larger datasets to achieve worse results: Context-1 utilized over 17,200 training items, while Search-R1 relied on a staggering 221,300 items to learn search behaviors.

By proving that a smarter external cognitive architecture can replace brute-force data scaling, Harness-1 suggests that the future of agentic AI lies in building better environments for models to work within, rather than just training larger models on more data.

Product: Enterprise Applicability and Generalization

From a product perspective, Harness-1 is delivered as a highly capable 20B agent merged into the openai/gpt-oss-20b base architecture.

For enterprise tech stacks, the applicability is massive because businesses need AI to execute multi-step research across proprietary databases without hallucinating or running up exorbitant compute bills.

Harness-1 manages its frontier-level performance at what the creators describe as “Context-1-level cost and latency.” Because the context window is strictly managed by the budget-aware harness rather than continuously expanding, enterprises can deploy this agent autonomously without incurring the exponential token costs typically associated with long-horizon AI tasks.

Even more impressively, Harness-1 proves it can generalize well beyond its training data. According to the research team, it was incredibly cheap to train, utilizing just 899 filtered supervised fine-tuning (SFT) trajectories and a mere 3,453 reinforcement learning (RL) queries.

“Instead of training the model to survive a giant append-only transcript, we train it to use a structured search interface: search, curate, revisit, verify, and submit,” Jiang explained.

This leanness proves a critical point for the AI industry: developers do not necessarily need petabytes of new behavioral data if they build a better cognitive framework for the model to operate within.

Licensing: The Power of Apache 2.0

One of the most significant aspects of the Harness-1 release is its licensing. In plain language, Apache 2.0 is a highly permissive, enterprise-friendly software license that fundamentally enables commercialization.

Unlike “copyleft” licenses (such as the GPL) that can force companies to open-source their own proprietary software if they integrate the code, or “research-only” licenses that ban commercial use entirely, Apache 2.0 gives businesses the green light to freely build, modify, and monetize the technology.

For developers and startups, this means Harness-1 can be seamlessly integrated into commercial enterprise search products, internal data retrieval tools, or customer-facing AI applications without fear of legal reprisal.

The only major requirement is that users must include the original copyright notice and explicitly state any significant modifications they make to the source code, positioning Harness-1 as a highly viable foundational building block for the enterprise.

Community Reactions: A Resounding Validation

The announcement has clearly struck a nerve within the developer community, validating the very real pain points engineers face when building agentic systems. Jiang’s multi-part announcement thread on X quickly garnered massive traction, pulling in over 256.1K views, 3.7K likes, 2.9K bookmarks, and nearly 300 reposts within a matter of days.

This high engagement underscores a growing consensus in the AI space that brute-forcing context windows is a losing battle.

When Jiang posted on X, “I’ve been wondering: maybe search agents are bad at search partly because we make them do all the paperwork in their head,” the resonance was immediate.

For developers who have spent the last year wrestling with AI agents that confidently forget their primary instructions halfway through a database search, the Harness-1 approach feels like a desperately needed course correction.

Ultimately, the community sentiment highlights a shift in industry priorities. Developers are moving away from asking how large an AI model’s context window can get, and instead asking how efficiently an AI model’s environment can manage that context for it. By offloading the paperwork, Harness-1 is proving that smaller, smarter systems can outmaneuver the giants—provided they have the right desk to work at.