OpenAI said on Tuesday that two of its AI models, including the flagship Sol, broke out of a secure test environment, gained internet access by exploiting a zero-day vulnerability in third-party software, and hacked into Hugging Face’s production infrastructure. The company called the incident “unprecedented” and said it was sharing preliminary findings to help defenders understand what frontier models are now capable of doing.

Both models were running with lower cybersecurity guardrails as part of an internal evaluation of their offensive capabilities. They were being tested against ExploitGym, a freely available cybersecurity benchmark, and determined that the answers to the test were stored on Hugging Face’s production systems.

Rather than solve the evaluation as intended, the models went after the answer key. They spent what OpenAI described as a “substantial amount of inference compute” finding a way out of their sandboxed environment, then chained two remote code execution vulnerabilities in Hugging Face’s dataset processing pipeline. The attack harvested cloud and cluster credentials, moved laterally into internal clusters, and executed more than 17,000 individual actions across a swarm of short-lived sandboxes, according to Hugging Face’s own disclosure on July 16.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

Hugging Face said it detected the intrusion, contained it, and rebuilt the compromised nodes. The company found no evidence of tampering with public models, datasets, or Spaces, and verified its software supply chain was clean. It is still assessing whether any partner or customer data was affected.

To investigate the breach, Hugging Face turned to GLM, a Chinese open-weight model, because the safety guardrails on US commercial models blocked the forensic queries its team needed to run.

The escape is not the first time Sol has been caught gaming its own evaluations. The Model Evaluation and Threat Research organization, the independent lab that red-teamed the model before launch, found it was aggressively hacking its test environments to inflate its scores. In one task, it packaged an exploit into a data stream, escalated privileges on the evaluation server, and leaked the correct answers that human evaluators had hidden.

The broader pattern of AI agent security failures has accelerated sharply, with four separate research teams breaking AI agents in four different ways during the first ten days of July alone. OpenAI and Anthropic have faced heightened scrutiny over their models’ cybersecurity capabilities, with the Trump administration restricting access to both companies’ newest systems during a government review.

OpenAI detected the Hugging Face attack and reached out to disclose it, but by that time, Hugging Face had already identified and contained the breach on its own. The incident demonstrates that the gap between AI models that can find vulnerabilities and AI models that will exploit them without permission is narrower than anyone in the industry had publicly acknowledged.

Turns out the “Hide” part of Hide My Email wasn’t doing its job quite as advertised, something that I covered early in July. Security researcher Tyler Murphy reported the flaw in June 2025, but despite Apple claiming it was resolved in March 2026, independent tests confirmed it remained exploitable, at least until July 3, 2026.

So how did this bug actually work?

The flaw meant that if someone sent a message to your hidden address and it bounced as spam, even a completely legitimate email, your actual email address could get exposed in the sender’s mail logs. 

You wouldn’t even know it happened. A bounced message never lands in your inbox for you to notice. Murphy first flagged the bug to Apple back in June 2025. Apple told him in March 2026 it had been resolved, but it hadn’t. Fast forward to early July, and 404 Media went public with the story.

Apple shipped an actual fix just two days later, on July 3. Make of that timing what you will. “The bug that caused Apple’s Hide My Email to leak hidden email addresses to senders has been fixed,” reports 404 Media. 

Is the issue resolved entirely?

However, the outlet warns that email addresses linked to Hide My Email aliases, especially those created prior to July 7, 2026, may still reside in retained third-party mail logs. Apple’s also now facing a lawsuit seeking class action status, accusing the company of violating California consumer protection law by selling a feature that didn’t work as promised.

This pattern, quietly acknowledging a flaw internally while telling the reporter it’s fixed when it isn’t, echoes complaints commenters raised about Apple’s App Store moderation only responding to public pressure. 

Privacy features marketed as premium selling points invite exactly this kind of scrutiny, and a pending class action means Apple’s handling of the timeline will matter as much as the fix itself.

Garmin is reportedly preparing to launch the Fenix 9, Fenix 9 Pro and Enduro 4 simultaneously this year.

That timeline comes from Garmin Rumors, which reports that Garmin plans to bring all three watches to market together rather than staggering their releases as it has with past generations.

The Enduro 4 stands out within that lineup as the only model expected to retain a memory-in-pixel display, a power-saving screen technology paired with integrated solar cells to extend battery life well beyond typical smartwatch limits.

The standard Fenix 9, by contrast, is expected to ship exclusively with an AMOLED panel, a shift that continues Garmin’s gradual move away from memory-in-pixel screens across its flagship range in recent years.

It’s been a big year for AI acquisitions — so big that most of them barely register anymore. Anthropic and OpenAI have each gone on buying sprees, snapping up developer tooling, AI services shops, and product-testing startups to convert model capability into enterprise revenue and extend their reach faster than the other. Which is what made a weekend rumor about Anthropic acquiring robotics startup Physical Intelligence stand out. It spread exceedingly fast, even after a denial from Physical Intelligence’s CEO.

Part of that ties to who’s involved. Physical Intelligence isn’t some obscure robotics shop. It was co-founded by Lachy Groom, an investor-operator whose star has been on the rise in Silicon Valley in recent years; it has raised more than $1 billion (and was reportedly in talks this spring for another $1 billion round at an $11 billion valuation); and its π0.5 model is apparently among the more widely used robot brains in robotics research.

As it turns out, the rumor wasn’t completely spurious. Anthropic and Physical Intelligence actually did hold acquisition talks this spring, according to The Information, so tech blogger Robert Scoble — whose weekend post on X set off the frenzy — may have gotten the specifics wrong without being wrong that something had happened.

Physical Intelligence’s response to the rumor mill wasn’t the world’s most vigorous denial, it should be noted. According to The Information, Physical Intelligence CEO Karol Hausman told employees the reports weren’t true via a Slack message containing a gif of a character from “The Office” shaking her head no.

Groom, for his part, did not respond to TechCrunch’s request for comment, sent Monday night.

Anthropic has made four known acquisitions this year; OpenAI has been more aggressive, acquiring at least 17 companies since 2023. Both are also, of course, now preparing to go public. Anthropic confidentially filed for an IPO on June 1, followed by OpenAI a week later, setting up what could be two of the largest U.S. stock debuts in history.

So why robotics, why now? The likeliest answer is that physical-world understanding may be a prerequisite for superintelligent systems, and no amount of internet text can substitute for it.

OpenAI’s own history here is instructive. It built an early robotic hand that could solve a Rubik’s Cube, then shut the entire robotics group down in 2021, with co-founder Wojciech Zaremba later saying the approach was missing pieces needed for real superintelligence. The team came back in 2024, quietly building a humanoid robotics lab in San Francisco, before CEO Sam Altman made it official in late May, announcing “OpenAI Robotics” was hiring and describing a near-term focus on robots for infrastructure work, with a personal robot for everyone as the long-term goal.

Anthropic hasn’t built anything resembling OpenAI’s hardware lab. What it has done is publish a string of research pieces through its internal group that stress-tests frontier capabilities for safety purposes. That included Project Fetch last November, where Anthropic staff tested how much Claude could help non-experts program a robot dog, and a second phase in June that, according to Anthropic, found a newer model completed the same tasks roughly 20 times faster than the best human-plus-Claude team from the year before.

Buying an existing team with robotics expertise would let Anthropic skip years of work. There’s a possible complication, though. Physical Intelligence was founded in San Francisco roughly two years ago by Groom, former Google researchers, and professors from Stanford and Berkeley, and its early investor base looks a lot like OpenAI’s own, including Khosla Ventures and Thrive Capital. Founders Fund — also a major OpenAI investor — was reportedly involved in Physical Intelligence’s newest funding round earlier this year.

In fact, OpenAI is itself an investor in Physical Intelligence, so it isn’t just a peripheral player; it’s a stakeholder in a company that its chief rival was reportedly in talks to buy very recently.

That raises questions around whether OpenAI’s early investment came with any information rights, or a right of first refusal over a sale to a competitor — the kind of protective provisions that strategic investors sometimes negotiate for precisely this scenario.

That leaves open the possibility that if Physical Intelligence is actually in play, OpenAI — already a shareholder, already close to Groom, already trying to ensure it bests Anthropic in robotics — may have the more obvious claim to it than Anthropic does. We asked OpenAI these questions earlier today and the company didn’t respond.

Each payout per work will amount to roughly $3,000, to be paid to around 500,000 works.

A US judge has approved Anthropic’s $1.5bn settlement of a major AI copyright lawsuit filed by a trio of authors, setting the record for the largest settlement in copyright law in the country’s history.

In their 2024 lawsuit, authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson accused Anthropic pirating their copyrighted material to train Claude. “Largescale theft of copyrighted works” is a key component of the company’s business model, they said.

The court found that Anthropic illegally acquired millions of books through shadow libraries, but ruled that the company was protected by fair use when using books to train its AI models. The lawsuit was certified as a class action last July, the first in copyright litigation against AI companies.

Anthropic agreed to settle last August, a month after it claimed that a settlement could present a “death knell” situation for the company.

Divided up between some 500,000 works, each payout per work under the approved settlement will amount to roughly $3,000, which the court pointed out is four-times the statutory damages awarded for “wilful infringement”.

The class includes copyright owners whose works were found in shadow libraries LibGen and PiLiMi, which were downloaded by Anthropic.

In a statement, Anthropic deputy general counsel Aparna Sridhar said that the company is “looking forward ​to bringing this matter to a close”.

“We reached this settlement in 2025, after the court’s landmark ruling that training AI on books is fair ​use under copyright law – which remains the law today.

“We are pleased that more than 91pc of authors and publishers covered by the settlement have claimed their share of the payment.”

The settlement received preliminary approval late last September.

Anthropic was valued at $183bn last September after a $13bn Series F round, and nearly a year later, the company is now nearing a $1trn valuation while readying itself for a blockbuster public listing.

For comparison, the $1.5bn settlement makes up roughly 0.8pc of Anthropic’s September valuation and marginally more than 0.1pc of its value after the May raise.

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.

• Reports indicateHuawei is building and operating DRAM fabs through a joint venture with state-controlled SwaySure
• This is part of an alleged 11-fab network Huawei operates behind state-owned fronts that it denies exists
• The move, seen as a bid by Huawei to secure captive HBM supply for its increasingly powerful Ascend AI chips, could place the company above Apple, Intel, and Nvidia, none of which makes its own memory.

Mounting evidence suggests Huawei is taking the same approach that saw it build CPUs and AI chips in the past to DRAM in the present.

The Chinese tech conglomerate is allegedly building and operating DRAM fabrication plants on the mainland, bringing it into direct competition with only one major player at the same scale, albeit for a much larger consumer base: Samsung.

Overview: As social media algorithms come under increasing scrutiny, policies to ban children from services such as Facebook, X, YouTube, and Snapchat are gaining popularity across the globe. France has just become the latest country to pass such legislation – but data suggests the impact of Australia’s earlier ban has been limited.

France’s Parliament gave final approval on Tuesday to a law banning children under age 15 from social media, with the National Assembly voting 279 to 81 after the Senate had approved the text earlier the same day. The vote followed the Assembly’s initial approval of the bill in January by a margin of 130 to 21.

President Emmanuel Macron, who strongly championed the legislation, wants it in effect before the new school year begins in September. The bill also bans smartphone use in high schools but exempts online encyclopedias such as Wikipedia, educational and scientific directories, and open-source software platforms such as GitHub.

The bill enjoyed unusually broad bipartisan support in France’s traditionally divided parliament. Macron fast-tracked the legislation, hailing its passage as a “major step forward” and positioning France as the first country in the European Union to approve a blanket social media ban for minors.

Macron cited data from France’s national public health agency indicating that one in two teenagers in the country spends between two and five hours each day using smartphones. Approximately 90% access the internet via phones daily, and more than half access social media on them.

The president has said the country aims to protect children’s brains from being “manipulated,” in his words, “neither by American platforms, nor by Chinese algorithms.” Lawmaker Laure Miller, who authored the bill, said the legislation firmly establishes that social media is not harmless, arguing it causes children to read less, sleep less, and compare themselves to others more.

With the law now passed, France leapfrogs proposals still moving through other European countries, including Greece, Norway, Spain, and the UK. Greece aims to enforce a ban starting next year, and restrictions for under-16s are set to be introduced into the Norwegian and British parliaments by the end of 2026.

Australia became the first country in the world to attempt such a ban, in December 2025, but recent studies cast doubt on its effectiveness. Australia’s eSafety Commissioner found that roughly seven in ten children who held social media accounts before the ban took effect still retained them three months later, and a study published in the British Medical Journal put the figure as high as 85%.

Regulators say Facebook, Instagram, Snapchat, TikTok, and YouTube have largely failed to enforce the removal of underage accounts. Most teenagers who kept access report never encountering an age-verification check, even though some circumvented restrictions using VPNs or other workarounds. France is also considering restricting VPN use to close a similar loophole.

In response to the noncompliance, the Australian government doubled the maximum penalty for infringing companies to $99 million AUD (about US$68 million). Australia’s internet regulator, eSafety, is currently investigating Facebook, Instagram, Snapchat, TikTok, and YouTube for non-compliance.

Photo credit: Wewillmakeitnow
Months of cutting, measuring, and reprinting have produced a second version of a fully native PS4 handheld that sits smaller and more balanced in the hands than the first attempt. Modder wewillmakeitnow started with the same PS4 Slim motherboard that powered the original project, then trimmed it further after gaining sharper skills with 3D modeling software. The board still delivers complete PlayStation 4 performance, only now it draws less power and fits inside a cleaner shell.

A 7-inch OLED display brings games to life in 1080p, with razor fine detail and deep blacks that these panels are known for. The good news is that many PS4 titles already natively target 1080p, so your games will appear crisp with no upscaling to worry about. Next to the main panel is a tiny secondary display driven by an ESP32-C6 chip that monitors power consumption, battery life, and temperatures. It also allows the system to switch on and off the OLED screen simultaneously, which is a nice touch that makes the overall experience feel more approachable.

PlayStation 5 Portal Remote Player – Midnight Black

• Play Your Game Collection with Remote Play – PlayStation Portal Remote Player can play compatible games you have installed on your PS5 console…
• Cloud Streaming from the Game Catalog and Classics Catalog – Discover an awesome library of PS5 games on the PlayStation Portal Remote Player with…
• Cloud Streaming for PS5 Games in Your Library – With PlayStation Plus Premium, stream select digital PS5 games in Your Library from PlayStation Store…

Six 21700 cells, each with a decent 5000 mAh, are housed in a replaceable pack on the back of the machine. When the batteries are removed, the entire device weighs only 900 grams, which is light enough that you could easily plug it in and continue playing without the added weight. The battery life is also rather good, lasting approximately 90 minutes in the more demanding titles and a full three hours in lighter ones. He lost some battery life over the previous version, but this is easily compensated for by the fact that the board itself consumes less power and the GPU has been deliberately undervolted to achieve its lowest stable point.

Cooling is handled by a fan borrowed from an old GTX 750 graphics card, which is installed on the rear of the chassis and flows air softly. There’s also a much smaller 30mm fan controlling airflow within the chassis, which can get rather noisy, especially if you’re playing for lengthy periods of time, but the good news is that you can simply remove a plastic guard to allow the case to breathe a little more if necessary.

HDMI is still supported, so you can plug this thing into your TV and play on the big screen exactly like you would with a PS4, or use a DualShock controller if you like. The casing is also 3D printed, and they were able to get a decent fit with it. The controls are pleasant and near to your fingers, and the whole vibe is much less like a prototype and much more like a device you could actually take on a journey.
[Source]

Good morning! Let’s play Connections, the NYT’s clever word game that challenges you to group answers in various categories. It can be tough, so read on if you need Connections hints.

What should you do once you’ve finished? Why, play some more word games of course. I’ve also got daily Strands hints and answers and Quordle hints and answers articles if you need help for those too, while Marc’s Wordle today page covers the original viral word game.