OpenAI is exploring possible legal action after users did what they wanted with Apple’s ChatGPT integrations and didn’t sign up for enough paid accounts, instead of doing what CEO Sam Altman expected.

A May 14 Bloomberg report says OpenAI has enlisted outside legal counsel and discussed options that could include sending Apple a breach-of-contract notice. OpenAI reportedly expected deeper ChatGPT integration across iOS, iPadOS, and macOS to drive large subscription growth through Apple’s ecosystem.

Executives at OpenAI now believe the partnership has been financially disappointing and far more limited than anticipated. Apple and OpenAI entered the agreement with distinct priorities.

Apple needed a recognizable AI partner while major Siri upgrades and Apple Intelligence features were still under development. OpenAI sought access to hundreds of millions of Apple users, believing the iPhone could become a significant source of recurring ChatGPT subscriptions worth billions annually.

Internally, the companies reportedly viewed the OpenAI agreement as a potential counterpart to Apple’s enormously profitable Google Search deal in Safari. The ChatGPT partnership never came close to generating that level of revenue or strategic value.

Neither company structured the partnership around large direct payments. Instead, both sides expected strategic benefits. OpenAI expected customer growth and subscriptions, while Apple gained an interim AI solution as it continued building its own generative AI systems.

Apple kept ChatGPT available, but tightly controlled

Despite Apple’s high-profile WWDC 2024 announcement, ChatGPT inside Apple’s software exposes fewer features than OpenAI’s standalone app and operates within much tighter limits.

Users often need to explicitly invoke “ChatGPT” inside Siri prompts before requests route to OpenAI’s systems. Responses also appear inside smaller interface windows with less information than the standalone app typically provides.

OpenAI’s internal studies found users overwhelmingly preferred the standalone ChatGPT app over Apple’s built-in integrations.

Apple’s version resembles a tightly managed Siri extension rather than a deeply integrated AI layer across the operating system. The standalone app offers features unavailable in Apple’s implementation, such as persistent memory, wider model access, advanced voice tools, custom GPTs, and direct subscription management.

OpenAI executives reportedly expected broader integration across Apple apps and more prominent placement inside Siri. Apple instead kept the implementation relatively narrow while continuing development of its own Apple Intelligence systems.

Apple reportedly worried internally about OpenAI’s privacy standards while negotiating the partnership. The company built Apple Intelligence around on-device processing and Private Cloud Compute to keep more user data inside Apple-controlled infrastructure.

Apple Intelligence

That tension over data control goes to the heart of why the two companies were never fully aligned. OpenAI’s cloud-focused systems operate very differently, leaving Apple with less direct control over how outside AI models process user information.

OpenAI increasingly looks like a future Apple competitor

The relationship between the companies has also become more complicated outside software. OpenAI acquired Jony Ive’s AI hardware startup and has aggressively recruited Apple engineers for its growing device ambitions.

OpenAI offered some Apple recruits compensation packages worth millions more than Apple provided. Those moves increasingly position OpenAI as a competitor rather than a partner.

Apple is also building new hardware efforts tied to Jony Ive‘s AI startup while preparing for a future where outside AI providers become interchangeable services inside iOS.

Apple plans to introduce an “Extensions” system in iOS 27 that will let users choose among multiple outside AI models. ChatGPT, Anthropic’s Claude, and Google Gemini are all expected to be part of the system.

Outside AI models are becoming interchangeable services inside iOS rather than central platform features. ChatGPT appears set to operate alongside Siri and Apple Intelligence instead of becoming a dominant layer across Apple’s ecosystem.

Any legal case could still prove difficult. Large platform agreements often give companies like Apple broad control over implementation details, placement, and product design decisions. And, users will do what they want, regardless of Silicon Valley expectations.

OpenAI still hopes to resolve the disagreement privately and may wait until after its legal fight with Elon Musk concludes before taking formal action against Apple. A lawsuit isn’t guaranteed at this stage.

A code migration agent finishes its run, and the pipeline looks green. But several pieces were never compiled — and it took days to catch. That’s not a model failure; that’s an agent deciding it was done before it actually was.

Many enterprises are now seeing that production AI agent pipelines fail not because of the models’ abilities but because the model behind the agent decides to stop. Several methods to prevent premature task exits are now available from LangChain, Google and OpenAI, though these often rely on separate evaluation systems. The newest method comes from Anthropic: /goals on Claude Code, which formally separates task execution and task evaluation.

Coding agents work in a loop: they read files, run commands, edit code and then check whether the task is done. 

Claude Code /goals essentially adds a second layer to that loop. After a user defines a goal, Claude will continue to turn by turn, but an evaluator model comes in after every step to review and decide if the goal has been achieved. 

The two model split

Orchestration platforms from all three vendors identified the same roadblock. But the way they approach these is different. OpenAI leaves the loop alone and lets the model decide when it’s done, but does let users tag on their own evaluators. For LangGraph and Google’s Agent Development Kit, independent evaluation is possible, but requires developers to define the critic node, write up the termination logic and configure observability. 

Claude Code /goals sets the independent evaluator’s default, whether the user wants it to run longer or shorter. Basically, the developer sets the goal completion condition via a prompt. For example, /goal all tests in test/auth pass, and the lint step is clean. Claude Code then runs, and every time the agent attempts to end its work, the evaluation model, which is Haiku by default, will check against the condition loop. If the condition is not met, the agent keeps running. If the condition is met, then it logs the achieved condition to the agent conversation transcript and clears the goal. There are only two decisions the evaluator makes, which is why the smaller Haiku model works well, whether it’s done or not. 

Claude Code makes this possible by separating the model that attempts to complete a task from the evaluator model that ensures the task is actually completed. This prevents the agent from mixing up what it’s already accomplished with what still needs to be done. With this method, Anthropic noted there’s no need for a third-party observability platform — though enterprises are free to continue using one alongside Claude Code — no need for a custom log, and less reliance on post-mortem reconstruction.

Competitors like Google ADK support similar evaluation patterns. Google ADK deploys a LoopAgent, but developers have to architect that logic.

In its documentation, Anthropic said the most successful conditions usually have: 

• One measurable end state: a test result, a build exit code, a file count, an empty queue

• A stated check: how Claude should prove it, such as “npm test exits 0” or “git status is clean.”

• Constraints that matter: anything that must not change on the way there, such as “no other test file is modified”

Reliability in the loop

For enterprises already managing sprawling tool stacks, the appeal is a native evaluator that doesn’t add another system to maintain.

This is part of a broader trend in the agentic space, especially as the possibility of stateful, long-running and self-learning agents becomes more of a reality. Evaluator models, verification systems and other independent adjudication systems are starting to show up in reasoning systems and, in some cases, in coding agents like Devin or SWE-agent. 

Sean Brownell, solutions director at Sprinklr, told VentureBeat in an email that there is interest in this kind of loop, where the task and judge are separate, but he feels there is nothing unique about Anthropic’s approach.

“Yes, the loop works. Separating the builder from the judge is sound design because, fundamentally, you can’t trust a model to judge its own homework. The model doing the work is the worst judge of whether it’s done,” Brownell said. “That being said, Anthropic isn’t first to market. The most interesting story here is that two of the world’s biggest AI labs shipped the same command just days apart, but each of them reached entirely different conclusions about who gets to declare ‘done.’”

Brownell said the loop works best “for deterministic work with a verifiable end-state like migrations, fixing broken test suites, clearing a backlog,” but for more nuanced tasks or those needing design judgment, a human making that decision is far more important.

Bringing that evaluator/task split to the agent-loop level shows that companies like Anthropic are pushing agents and orchestration further toward a more auditable, observable system.

OpenAI is so frustrated with Apple over a ChatGPT integration that failed to deliver the subscribers and prominence it expected that the company is now actively exploring legal action against the iPhone maker, Bloomberg News reported Thursday, citing people familiar with the matter.

According to Bloomberg, OpenAI has enlisted an outside law firm to work through its options, which could include sending Apple a formal breach-of-contract notice without necessarily escalating to a full lawsuit (at least not immediately). Any legal move would likely wait until after the conclusion of OpenAI’s ongoing trial with Elon Musk.

Still, it’s a reminder of what a difficult partner Apple can be for major software companies. The iPhone is an enormously attractive platform for growth, but it’s fully under Apple’s control — and companies that build there are only guests. From Google to Adobe, there’s a long history of Apple showing guests the door when they seem as if they’re getting too comfortable.

TechCrunch has reached out to both OpenAI and Apple for comment.

The OpenAI partnership, announced at Apple’s Worldwide Developers Conference in June 2024, wove ChatGPT into Apple’s operating systems as an option within Siri and as part of the iPhone’s Visual Intelligence feature (allowing users to use their camera to analyze their surrounds and send photos to ChatGPT with related questions).

OpenAI, along with industry watchers, expected the deal might eventually funnel billions of dollars in new subscriptions its way and give the company prime real estate across one of the world’s most-used mobile ecosystems. Instead, Bloomberg reports, OpenAI has grown increasingly aggravated, complaining that the integration has been buried, its features hard to find, and that revenue from the tie-up is nowhere close to projections. “They basically said, ‘OpenAI needs to take a leap of faith and trust us,’” one OpenAI executive told Bloomberg. “It didn’t work out well.”

Apple, for its part, has its own grievances, including concerns about OpenAI’s privacy standards and, according to Bloomberg, irritation over OpenAI’s push into hardware, an effort led by former Apple executives including ex-design chief Jony Ive.

Either way, OpenAI is hardly the first partner of Apple to regret hitching its wagon to the company. Apple has a long history of embracing partners and then alienating them. The most famous case is Google Maps, which was a flagship feature of the original iPhone. It was so central to the device’s appeal that its removal in 2012 — replaced by Apple’s markedly inferior Apple Maps product — became one of the biggest tech fiascos of the decade, prompting a rare public apology from CEO Tim Cook. The friction between the two companies had been building for years at that point, thanks to the rollout of Google’s Android phone a year after the iPhone’s 2007 debut; after Google’s then-CEO Eric Schmidt stepped down from Apple’s board in 2009, that rivalry only intensified.

Adobe has some scar tissue, too. Steve Jobs refused to support Flash on the iPhone and iPad, publishing a famous open letter in 2010 explaining why and effectively dooming the technology. Flash never recovered its footing on mobile.

Then there’s Spotify, which spent years arguing that Apple leveraged its control over the App Store to disadvantage rival music streaming services after launching Apple Music in 2015. The European Commission agreed, fining Apple nearly €1.8 billion in March 2024.

Sometimes these rifts can be overcome in the name of commercial interests. Google is now Apple’s AI infrastructure partner, having struck a multiyear deal in January to power the next generation of Apple Intelligence with Gemini models. Apple is paying Google roughly $1 billion a year.

In the meantime, OpenAI has had its own share of strained relationships lately. Elon Musk’s lawsuit against the company — which accuses OpenAI of abandoning its nonprofit founding mission and operating in bad faith — is currently at trial.

The company has also reportedly navigated tensions with Microsoft, its biggest backer and infrastructure partner, as it pushes for greater independence ahead of its own IPO ambitions.

The TWS earbuds category, at least in India, is a tricky business. There are too many players, each with their specific appeal. Noise is one such homegrown brand that’s always appealed to budget-conscious buyers. I was one of the early adopters when they first started doing smart bands. Their earphones have been solid overall, but have struggled to stand out among more established players like OPPO and OnePlus. On the flip side, Bose is the most recognized player in the premium headphone space, as they essentially invented ANC back in the 90s.

So, what if you combine the value proposition of Noise with the premiumness and sound of Bose? You get the Noise Master Buds. The first generation of these unique collaboration earbuds was a hit, with experts praising the sound quality and the ANC capabilities. Fast forward to today, and the second generation of the Master Buds are out, bringing some meaningful updates in the sound department. And I called Noise to get these for a review. It’s been a month since that call, and I even took them with me to Thailand. Spoiler alert: these earbuds are a hit. Here’s why.

Design & Comfort

When I first got the Master Buds 2, I was genuinely in awe of the design. I genuinely haven’t seen a case this pretty before, yet still functional. The concentric circle pattern on the front is a really nice touch that adds character to the stale world of TWS earbuds. The semicircle shape is also unique and something I have yet to see from other makers. One more benefit of this shape is that you can keep the case standing vertically, almost resembling the clocks of the past, or maybe I’m just too old at this point.

Nevertheless, I was constantly asked by my friends and family what earbuds I was using, all of whom adored the design, so you know it’s not just me yapping. I’d recommend sticking to the silver color, as it’s the best and doesn’t pick up random scratches, but the black also looks pretty decent. There’s a diagonal white LED strip that’s meant to indicate when the earbuds are on. You also get a dedicated pairing button, so no hand gymnastics is needed. Still, the best part of the design for me is the opening/closing mechanism. There’s a good amount of weight behind it, which makes that heavy thud that I love.

I do, however, have a problem with the size. I know these earbuds pack a lot of hardware, but the case size is just massive. I’ve been daily driving the Enco Buds 3 Pro+ for months, and compared to them, the Master Buds feel huge.

Moving to the earbuds themselves, comfort is highly subjective, as everyone’s ears are different. All that being said, though, the first day of me using the Master Buds wasn’t the best. My ears are small, and the medium-sized tips that come pre-installed were just too big for my liking. Thankfully, when I put on the small eartips, the experience was much better. You’ll still feel them sticking out, but the rubber fins work well to keep them from falling out during my everyday gym struggles with weights. It’s been a couple of hours since I started writing this review, and the Master Buds 2 are still sitting comfortably in my ears.

Sound Quality & Battery Life

Often, budget earbuds lean heavily towards bass and compromise everything else. Well, that’s not the case with the Master Buds 2. The 10 mm drivers tuned by Bose prioritize balance over everything else, and I love that. The mids, where most of the vocals live, are crisp, with you being able to hear the small nuances of a singer’s voice. The highs are decent without that sharpness that pinches through, and I love the separation between the different elements. Everything is placed perfectly, and that’s what many earbuds fail to do. Bass isn’t the selling point of the Master Buds, but the lows are still there, working hard in the background. Just don’t expect a rumble.

If you’re not happy with the signature Bose tuning, which would be surprising, Noise does bundle a few different listening modes like Jazz, Club, and Rock, each with its unique style. For audiophiles, there’s also a custom equalizer. Beyond the basics, Noise has bundled spatial audio with head tracking. I tried some songs on Apple Music, and it works great with instruments placed around you that change as you move your head. Call quality was excellent, with the other person hearing me loud and clear. Speaking of loud, though, there’s a new feature called Sidetone that lets you hear yourself on calls to check if you’re talking too loudly. Honestly, I can think of a lot of people needing that feature.

The Master Buds 2 use their 6 microphones for up to 51dB of active noise cancellation, and I put that number to the test. Using the ANC at Max, the buds do a stellar job of quietening the everyday hum of an AC or fan without a hitch. Voices were muffled enough that I couldn’t hear someone talking near me at half volume. For commuters, the Master Buds 2 will be enough to block out the chatter of a metro, but sudden loud noises, like a horn, will still find their way through. Last but not least, the battery on the buds is very decent. I got about 4-5 hours of use on a single charge. The case can recharge the buds at least 3 times, and once you do run out, fast charging should come to the rescue.

Companion App

It’s no secret that controls are super easy to mess up. Fortunately, that’s not the case with the Noise Master Buds. In fact, there’s a lot going on. Everything is controlled via the Noise Audio app, and it looks really well designed. It’s sophisticated, and everything is laid out neatly. Still, there are plenty of ways you can control music.

The obvious choice is the touch controls on the earbuds themselves. You can configure a single touch, a double touch, a triple touch, a quadruple touch, and even a tap-and-hold. But that’s not it. Noise has incorporated motion controls, meaning if you nod your head twice, you can play/pause the music. Beyond that, shaking your head twice to the right will skip to the next track and vice versa. This is a pretty niche feature that I thought would be a gimmick at best. Surprisingly, it works quite well, except for the occasional misses. You can also nod to accept calls.

No review of a 2026 product will be complete without mentioning AI, and the same is true here. Noise has bundled something called Noise AI, which, in theory, is meant to answer your questions. I gave it a go with one of the recommended questions: suggesting a good cafe near me. The answer was that it doesn’t have access to live restaurant listings and that I should search Google instead. If I have to search Google, then what’s even the point?

Verdict

At ₹8,999, the Noise Master Buds 2 sit in a very competitive segment, but they do enough differently to stand out. The design alone is a conversation starter, and unlike many flashy earbuds, these actually back it up with substance. The Bose-tuned sound signature focuses on balance rather than overwhelming bass, making everything from vocals to instruments sound clean and detailed. ANC performance is strong enough for everyday commutes and flights, the gesture controls are surprisingly useful, and the companion app is genuinely well-designed rather than feeling like an afterthought.

Of course, they aren’t perfect. The case is noticeably larger than most competitors, and bass lovers might find the tuning a little too restrained. Noise AI also feels half-baked right now and adds very little to the experience. Still, if you’re shopping in this segment, the Master Buds 2 are a must-consider.

The $200 discount applies to the M5 model with 1TB of storage, matching the lowest price on record. Plus, grab an even larger markdown on an upgraded spec.

The $1,499.99 special is available on the 14-inch MacBook Pro with a 10-core CPU and 10-core GPU. It also has 16GB of memory and 1TB of storage, with the $200 discount available in your choice of Space Black or Silver.

Buy M5/16GB/1TB MacBook Pro for $1,499.99

If you’d prefer additional RAM, B&H is running a $300 discount on the M5 model with 24GB of RAM. Note that this configuration comes with 512GB of storage, rather than 1TB, but it rings in at the same $1,499 price as the above deal when ordered in Silver.

Buy M5/24GB/512GB MacBook Pro for $1,499

Both of these deals can be found in our M5 MacBook Pro Price Guide, with markdowns on higher-end M5 Pro and M5 Max models available in our M5 Pro 14-inch MacBook Pro Price Guide and M5 Pro 16-inch MacBook Pro Price Guide.

Laden with iconic Royal Oak design cues, most notably the octagonal case, eight-screw bezel, and Petite Tapisserie-patterned dial, the strapless design heavily references 1979’s Royal Oak Pocket Watch reference 5691. Inside is an entirely new hand-wound version of Swatch’s Sistem51 caliber, a movement that is completely machine assembled. Swatch has 15 active patents on this new iteration and has also squeezed in an impressive 90-hour power reserve. There’s even an antimagnetic Nivachron balance spring that was, incidentally, codeveloped with Audemars Piguet.

Swatch’s 1986 POP line, whose watch heads could be physically ejected from their frames and clipped elsewhere, has been plundered here to create a design that allows the Royal Pops to ping out of their bioceramic holder clips, too.

Why There’s No Wristwatch

The simple logic of the pocket watch design authorized by Audemars Piguet, which, unlike Omega, is not part of the Swatch Group, is that it doesn’t upset its existing high-net-worth customer base. Royal Oak owners will no doubt be breathing sighs of relief now that it’s confirmed a version of their coveted pieces won’t be coming to market for a mere few hundred bucks.

However, this doesn’t mean that AP would have been financially hit had it delivered what the public so clearly wanted. Omega, which was also concerned for its sales when shown the original MoonSwatch internal prototypes, enjoyed a sizable 50 percent bump in sales following the release of its budget cousin.

The Royal Pop pocket watch, cleverly, is a sidestep designed to generate as much hype as possible yet be as safe as can be for AP’s brand. The Royal Oak design language is unmistakable, but the wrist is off-limits. With Swatch, Audemars built something real for its aspirational fans; it just didn’t build them what they wanted.

What does Swatch get out of this? Valuable PR as well, but far more importantly, the potential of a much-needed sales hit. In 2025, the group posted a 6.75 percent drop in sales and a staggering 55.6 percent decline in operating profit, primarily attributed to a sharp drop in demand for its watches in China, Hong Kong, and Macau. Swatch Group shareholders are not happy.

How China Will Come to the Rescue

Here is where the story gets interesting for reasons neither Swatch nor AP planned. As Swatch resurrected its POP design, allowing the Royal Pop to be removed from its housing, within hours of the Royal Pop announcement, third-party strap brands seized on this prospect, looking to quickly fashion adaptations that convert the timepiece from pocket to wristwatch. As Royal Pops were designed to snap in and out of lanyards and desk stands, they should just as easily clip into bracelets and straps made specifically to receive them.

The market recognized in real time that the pocket watch from Swatch and AP tantalizingly contained all that was structurally needed to deliver the very wristwatch that the AI concepts had promised. All that was required now was something to connect the case to a wrist.

The name is the tell. Graphon AI, which emerged from stealth on Wednesday with $8.3 million in seed funding, is named after a mathematical object that most people in AI have never heard of and that its two most prominent advisors helped invent. A graphon is the limit of a sequence of dense graphs: a continuous function that captures the structure of relationships as networks grow infinitely large. It is the kind of concept that exists at the boundary between pure mathematics and theoretical computer science, and it is now the foundation of a startup that claims to have built the missing layer between enterprise data and the models that are supposed to make sense of it.

The company’s thesis is straightforward, even if the mathematics behind it are not. Today’s large language models can process roughly one million tokens at a time. Enterprises hold trillions of tokens across documents, video, audio, images, logs, and databases. Retrieval-augmented generation, the current standard approach, can surface relevant content from that mass, but it cannot discover relationships between pieces of data that were never stored together. An LLM using RAG can answer a question about a specific document. It cannot reason about how that document connects to a surveillance video, a compliance log, and a customer database, at least not without someone having already mapped those connections.

Graphon’s product sits before the model, not inside it. Using graphon functions, a mathematical framework that extends the academic concept into a software layer, the system ingests multimodal data and automatically discovers relational structure across it, producing what the company calls persistent relational memory. The result, in theory, is a representation of an organisation’s data that any foundation model or agent framework can query without being constrained by its context window.

The people behind the mathematics

The founding team comprises Arbaaz Khan as chief executive, Deepak Mishra as chief operating officer, and Clark Zhang as chief technology officer. The company says its broader team includes former researchers and engineers from Amazon, Meta, Google, Apple, NVIDIA, Samsung AI Center, MIT, Rivian, and NASA.

More notable, perhaps, are the technical advisors. Jennifer Chayes, dean of the College of Computing, Data Science, and Society at UC Berkeley, and Christian Borgs, a UC Berkeley computer science professor, are both listed as advisors. Borgs was among the group of researchers, alongside Chayes, László Lovász, Vera Sós, and Katalin Vesztergombi — who formalised the graphon as a mathematical concept in 2008. The company is, in effect, commercialising a framework that its advisors co-invented.

Chayes and Borgs described the approach in a joint statement as one that treats relational structure as a first-class element of the AI stack rather than something to be inferred after the fact. The distinction matters because most current AI systems treat data as collections of individual items to be retrieved, not as networks of relationships to be preserved.

An unusual investor table

The seed round was led by Arvind Gupta of Novera Ventures, who made Graphon his fund’s first investment from its flagship vehicle. Gupta is better known as the founder of IndieBio, the life-sciences accelerator, and his pivot toward an AI infrastructure company suggests he sees structural overlap between the problems Graphon addresses and the complex, multimodal data challenges that define scientific computing.

The rest of the cap table reads like a deliberate exercise in strategic diversity. Perplexity Fund, the $50 million venture arm of the AI search company, participated alongside Samsung Next, Hitachi Ventures, GS Futures (the venture arm of South Korean conglomerate GS Group), Gaia Ventures, B37 Ventures, and Aurum Partners, the investment fund affiliated with the ownership group of the San Francisco 49ers.

The mix is telling. A search-AI company, a consumer electronics giant, a Japanese industrial conglomerate, and a Korean chaebol all investing in the same pre-model data layer suggests that the context-window problem Graphon claims to solve is felt across industries that otherwise have little in common. GS Group, which ranks among South Korea’s largest conglomerates with interests spanning energy, retail, and construction, is also an early customer. Ally Kim, a vice president at GS, said the company’s multimodal AI solutions have been applied to analysing customer movement in convenience stores and enhancing safety through CCTV analysis at construction sites.

The technical bet

Graphon’s positioning reflects a broader shift in the AI infrastructure market. The past three years have been dominated by a race to build larger models with longer context windows. But even the most capable models still hit a ceiling: they can process more tokens, but they cannot maintain relational awareness across the volumes of data that large organisations generate. The question Graphon is betting on is whether the solution lies not in extending the context window further, but in structuring data before it enters the window at all.

The company says it has already deployed its platform for enterprise content management, industrial intelligence, agentic workflows, and on-device applications across phones, cameras, wearables, and smart glasses. The breadth of claimed use cases is ambitious for a company at the seed stage, and the absence of independent benchmarks or detailed customer case studies beyond GS Group makes it difficult to assess how far the technology has progressed from concept to production.

What is clear is that the problem Graphon describes is real. The gap between what LLMs can theoretically do and what they can actually do with enterprise data remains one of the most significant constraints on AI deployment. Retrieval-augmented generation has been the industry’s primary answer, and its limitations, flat retrieval that misses cross-dataset relationships, context windows that force artificial boundaries on what the model can see, are well documented. Whether graphon functions offer a fundamentally better approach or merely a more theoretically elegant version of graph-based data structuring is the question the company will need to answer as it moves from stealth-mode mathematics to production-grade infrastructure.

The $8.3 million gives it runway to try. The advisors who co-invented the underlying mathematics give it credibility. But in an AI market that has seen no shortage of startups claiming to have found the missing layer, Graphon’s challenge will be proving that the mathematics it is named after translates into a measurable improvement in how foundation models handle the messy, multimodal reality of enterprise data, not just in theory, but at the scale where theory stops being sufficient.

Memorial DaY brings discounts to the mattress models we test all year long, and the sales have already started. As a seasoned deal hunter, I know that mattresses go on sale pretty often, but whenever someone asks me the best time to buy, I tell them to wait until Memorial Day or Black Friday and Cyber Monday. If you’ve been in the market for a new mattress, now’s the time to act.

The WIRED Reviews team thoroughly tests the best mattresses long-term. We don’t conduct “nap tests” or base recommendations on first impressions. Our top picks are tried-and-true, and they’re on sale right now. We’ll also include some deals on bedding, pillows, mattress toppers, and other sleep accessories as we update this story through Memorial Day on May 26. Prices shown are for queen sizes.

Feel free to check out our many other sleep recommendations, including the best pillows for neck pain, the best body pillows, and the best sunrise alarm cocks. You might also want to read our guide on how to choose a mattress.

WIRED Featured Deals:

Helix Sleep Midnight Luxe Hybrid for $1,824 ($675 off)—Use Code WIRED27

Use our exclusive coupon code WIRED27 to get 27 percent off our very favorite mattress for most people. We’ve seen it sell for about $100 less before, and they’ve thrown in more freebies, but this is still a great deal. Just be aware that the price might drop a little later in the month. In any case, the Midnight Luxe Hybrid is springy and medium-firm and should be well-suited to any style of sleeping. The individually wrapped springs are zoned so that you have more support where you need it to prevent back pain. It also doesn’t get too warm, though it’s thick enough that you’ll want deep-pocketed sheets. It’s been our favorite mattress for over eight years.

Security researchers have admitted that Anthropic’s Mythos AI model has been able to hack macOS, bypassing Apple’s security systems in a way never previously achieved.

Mythos is an early version of a new, more powerful Claude AI model software that is yet to be made public. Anthropic’s engineers have warned that it is too good at finding security exploits to allow it into the wild.

Now, proof of its abilities has come in the form of an escalation exploit. If used correctly, the exploit could potentially allow a hacker to gain control of a Mac despite Apple’s security measures.

Detailing the news, The Wall Street Journal says that the security researchers were “excited about their discovery.” In fact, they were so impressed with what Mythos had done that they drove to Apple’s Cupertino HQ to share their findings.

Chained attacks

The researchers, from a Palo Alto-based research outfit, say that Mythos didn’t use a single attack vector in its hack. Instead, it linked two bugs macOS together in an attempt to corrupt the Mac’s memory.

The macOS operating system has been hacked in a new way

Once the macOS memory had been compromised, Mythos was then able to “gain access to parts of the device that should be inaccessible.” It’s also possible that, should the hacks then be used alongside others, the Mac as a whole could become compromised.

For its part, a company spokesperson told the WSJ that it is reviewing and validating the security team’s findings.

“Security is our top priority, and we take reports of potential vulnerabilities very seriously,” Apple reportedly said. However, Apple hasn’t yet said whether it has patched the bugs Mythos used for its hack.

In fact, it isn’t clear what Mythos did and didn’t do right now. That shouldn’t be all that surprising, with details likely to remain fuzzy until Apple has addressed the security flaws that were leveraged.

However, the report also notes that the attack couldn’t be achieved by Mythos alone. Without the skills of the hackers working alongside the AI, it is believed the hack wouldn’t have been possible.

As for Mythos, Anthropic intends for it to be used for good. Project Glasswing was launched to allow Mythos to be used as a way to identify security flaws so they can be addressed.