The big picture: The RAM crisis has made the latest hardware so unaffordable that components from four years ago are storming sales charts. AMD and Nvidia have already responded to the demand by re-releasing chips from 2022, and now Intel is poised to follow suit.

Robert Hallock, Intel’s vice president and general manager of enthusiast channel business, recently told Tom’s Hardware that the company plans to continue offering Raptor Lake CPUs, first launched in 2022, for a while longer. Intel, like AMD, is responding to sustained demand for hardware that supports DDR4 RAM.

During a lengthy interview, Hallock noted a spike in demand for Alder Lake and Raptor Lake processors, which have recently become harder to find. While the exec did not mention specific products, he confirmed that the company aims to stabilize and maintain the supply of 10-nanometer chips that support Intel’s LGA 1700 socket “for years to come.”

The comments align with prior rumors that Intel might introduce another refresh for Raptor Lake – the lineup’s third – during the first half of 2027. Tom’s Hardware heard reports on the topic during Computex earlier this year, where at least two motherboard vendors also confirmed plans to increase production of aging LGA 1700-compatible boards to meet rising demand. Intel’s rumored refresh, tentatively dubbed Raptor Lake Next, would exist alongside the DDR5-exclusive Nova Lake.

Also Read: DDR5 vs. DDR4 Gaming Performance Revisited

Surging demand for processors and mainboards that support the supposedly obsolete DDR4 memory is likely due to RAM shortages driven by AI data centers. The price of DDR5 RAM has skyrocketed over the past year, driving PC users to either purchase DDR4 memory at much lower (but still inflated) prices or hold onto the RAM they already have. Raptor Lake CPUs support DDR4 and DDR5 memory, giving buyers the opportunity to upgrade without replacing an entire PC.

AMD has already responded to the situation by re-releasing the Ryzen 7 5800X3D, the CPU that introduced 3D V-Cache in 2022. The ongoing crisis has also turned the modest Ryzen 5 5500 into the reigning budget CPU king. While the company offers better processors at similar prices, such as the 7500F and 9800X3D, they all require DDR5 RAM.

The trend has even affected Nvidia GPUs. While tightening stocks of GDDR7 VRAM have increased the prices of the company’s latest RTX 5000-series graphics cards and delayed the launch of the RTX 5000 Super cards, Nvidia has revived the 12GB RTX 3060, also from 2022, as a temporary substitute.

Sam Altman told a summit of tech interns that two years of college was “the exact right amount of time” for him and that four years may be more than the world now requires. He made the case to an audience whose entry-level pipeline is already being closed by AI.

Sam Altman thinks a degree takes longer than it needs to. Two years at Stanford was “the exact right amount of time” for him, the OpenAI chief said, and staying would have brought “vastly diminishing returns.”

He was talking to tech interns. The remarks came during a surprise appearance at Internapalooza, a networking summit run by investor Cory Levy for students spending the summer in the industry.

“I’ve sort of thought that maybe the way the world has evolved, college just shouldn’t be as long as it is,” Altman said, while allowing that it was “still great to meet people and kind of like live on your own and get to work on projects.”

The audience is what complicates it. TNW has reported that AI is killing the summer internship, the mechanism that has long converted students into employees.

Nor is that only an American problem. A Swiss study found fewer job ads aimed at career starters as automation absorbed the tasks juniors used to be hired for.

Altman’s other advice pointed the same way without meaning to. “You can make a whole startup kind of by yourself in a room with a lot of AI tokens, but not much else,” he said.

Heard from a founder’s chair that is liberating. Heard from an internship, it is a description of a job that might not need filling.

Not everyone reads the trend that way. Adecco’s chief executive has argued that AI is changing jobs rather than destroying them, which is the more optimistic case and not an unreasonable one.

The rest of the session was conventional founder advice. People waste effort trying to be taken seriously, Altman said, though he conceded that enterprise sales was the exception and that he eventually solved it by hiring a 50-year-old.

His own route is worth remembering before anyone acts on this. Altman left Stanford in 2005 to co-found Loopt, which landed in Y Combinator’s first batch and led to him running the incubator, while his mentor Peter Thiel pays $250,000 over two years to under-22s on condition they drop out. As Altman put it himself, “you don’t get to run the experiment twice.“

What we know so far: Cybersecurity researchers have uncovered at least three new vulnerabilities that could potentially allow attackers to gain system privileges even on patched Windows computers. These include a memory configuration chip exploit named “Download more RAM,” a zero-day flaw dubbed “ShieldBreak,” and a “plug and pwn” exploit that could install malware remotely over RDP.

The “Download More RAM” vulnerability, which was presented at the 2026 USENIX Security Symposium in Baltimore, was discovered by researchers from the University of Birmingham and Durham University. It reportedly allows malicious actors to bypass Windows 11 security and gain system privileges without physical access by exploiting the lack of write protection on consumer memory modules.

The vulnerability allows anybody to remotely rewrite the Serial Presence Detect configuration chip that notifies the computer about the amount of installed RAM in the system. Attackers can take advantage of this vulnerability to transmit fake information to the computer, tricking it into believing that it has twice as much RAM installed as it actually does.

The false data tricks the memory controller into mapping additional pseudo-addresses, which overlap with genuine addresses. The aliases allow attackers to create a backdoor into memory by circumventing security and access control mechanisms used by the operating system and the processor.

The vulnerability could potentially allow hackers to re-enable old drivers with known exploits, disable anti-malware software, break into virtualization-based security enclaves, change corporate device management settings, and even bypass kernel-level anti-cheat systems used by video games.

Professor Tom Chothia from the University of Birmingham noted that the attack only requires a script that can be deployed remotely, while previous attacks of this kind required physical access to the machine. He added that major memory brands, including Corsair, G.Skill, and ADATA, ship at least one model line with an unprotected configuration chip in violation of JEDEC guidelines.

Tracked as CVE-2026-23670, the vulnerability has been acknowledged by both Microsoft and Corsair. While Microsoft issued mitigations in its April 2026 update, Corsair has added a feature to its iCue hardware management tool, allowing users to enable write protection on their DIMMs. PC diagnostic app HWiNFO has also added the same feature for non-Corsair users.

There’s also a zero-day vulnerability called ShieldBreak, found by bug hunter Nightmare Eclipse. Tracked as CVE-2026-50656, it’s an elevation-of-privilege vulnerability in Microsoft Defender that can circumvent the RoguePlanet patch and allow attackers to gain system privileges on Windows 10, Windows 11, and Windows Server. However, Microsoft Defender has to be active for the exploit to work.

Finally, security researchers Alejandro Hernando and Borja Martinez have described a new “plug and pwn” attack that exploits Windows’ automatic hardware identification and driver installation process to install signed vendor driver packages with system-level privileges.

Presented at DEF CON 34 in Las Vegas, the vulnerability can be exploited without admin privileges and without a logged-in user. In their proof-of-concept demo, the researchers also showed that the attack can be performed remotely over RDP without connecting any physical USB hardware to the target device.

CYBER-CRIME

Another one bites the dust

Some 1.6 million unique email addresses tied to RingCentral have been leaked online, alongside names, physical addresses, and phone numbers, according to Have I Been Pwned. RingCentral disclosed the breach on July 28 and said “it was the target of a sophisticated social engineering campaign” affecting a “limited portion of RingCentral customers.”

The comms platform said that it promptly responded to the intrusion upon detecting it, “took steps to stop the unauthorized activity,” and immediately launched an investigation into the security incident with help from a “leading third-party forensic firm.”

“We have not seen any new unauthorized activity since taking these remediation efforts,” the company added.

RingCentral did not immediately respond to The Register’s request for comment on this story. We will update it as needed. 

While the company hasn’t named its attacker, notorious data theft and extortion gang ShinyHunters previously claimed it compromised the collaboration platform, according to a post on its data leak site, viewed by The Register. Screenshots of the post also circulated on social media. The crooks claimed they stole more than 623 GB of data, and set a July 30 deadline for RingCentral to pay up – or else the crew would dump the stolen information online.

RingCentral apparently didn’t pay the extortion demand, and ShinyHunters followed through on its threat, posting customers’ details on the internet.

“The company failed to reach an agreement with us despite our incredible patience, all the chances and offers we made. They don’t care,” the crims wrote on August 3.

A ShinyHunters spokesperson told us that the group broke into RingCentral by voice-phishing an employee and tricking them into giving the crooks their password.

This same group, which security sleuth Dominic Alvieri says is his “top threat group and probably is for most analysts,” has hacked hundreds of organizations since the start of the year, including education tech firms that provide services for schools and universities along with healthcare-sector organizations. 

Recently, ShinyHunters dumped data stolen from Abbott’s cancer diagnostics business with the leak containing 10.9 million unique email addresses alongside personal and health information.

The crooks claim that they made off with more than 30 million rows of customer information, including more than one million Social Security numbers and 7.5 million dates of birth. More concerning, however, they said the haul includes 22 million-plus rows of client notes containing confidential doctor-patient conversations and health information, and more than 20 million medical-order records containing patient IDs, prescription types, order dates, and refill information.®

Editor’s note: This story was amended post-publication with comment from ShinyHunters.

After long evenings spent designing mechanisms and calibrating movement, a New Zealand engineer named Bruce finished a system that mounts directly onto an ordinary acoustic guitar and turns digital tablature into live string vibration. He calls it MegCell Pulse. Two coordinated robots handle the work of human hands. One presses the strings against the frets. The other plucks them with six individual plectrums. The sound that reaches the room is the actual instrument speaking, complete with the body resonance and the space around it.

Bruce began this project as an engineer rather than a musician, drawing on his skills in naval engineering, composite work, CAD drawings, and automation. In spare hours after his day job he kept asking what would happen if digital music instructions could drive a real guitar instead of a speaker. He spent roughly 6 or 7 years tinkering with the concept and its variants. The majority of the build is completed on a normal 3D printer, with the remaining parts sourced from a basic list that anybody can order. The software translates ordinary digital tablature into a precise timing signal for the actuators, and it does so without requiring special programming for each song.

Sale

Enya NOVA GO SP1 Carbon Fiber 35-Inch Smart Acoustic-Electric Travel Guitar

• WHY CARBON FIBER – Nova Go SP1 is made of carbon fiber composite material, which means it’s impervious to temperature and humidity change. Taking it…
• WHERE THINGS GET INTERESTING – The absolute differentiator of this guitar is that it’s equipped with the SP1 pickup system. It gets you reverb…
• A PRODUCTIVITY WORKHORSE – SP1 pickup system not only allows you to have tons of fun unplugged. But if you plug it into an amp or PA, this Nova Go SP…

After you put the frames onto the guitar, some gears and magnetic actuators begin moving the arms, which is quite cool. The fretting side simply presses each string in the proper location between the frets, while the plucking side strikes individual strings in sequence or all at once, depending on the file you imported. Timing remains fixed to the digital instructions, ensuring consistent performance from one run to the next. All the user has to do is load a tab file and hit play: the robots will do the rest, while the wooden body and strings generate the tone.

One of the most practical applications allows a human to handle only the plucking, leaving the fretting to the machine. A single-handed musician may complete a full arrangement this way, as Bruce has demonstrated in a video where the robot supplies the left-hand section while a person handles the right, which feels rather natural.

There are a few things it can’t do currently, such as not covering the entire fretboard or doing that nice sliding between frets for legato lines, but the limitations are real, and it still covers a large range of pieces if they are first available as digital tablature. The method focuses on accuracy and consistency rather than trying to record every single human flourish, which keeps the performances sounding rather clean.

No finished units are shipped from the factory, but Kickstarter backers receive the complete digital product, which includes STL files for the printed pieces, a detailed assembly guide, control software, and a parts list. The top tier was one hundred dollars. Once printed and assembled, the owner can simply fix or change it at home. That way, the project remains accessible while still allowing the design to grow and evolve in response to user feedback.
[Source]

If you want the best quality out of your Bluetooth audio, it’s worth reviewing some settings.

Understanding Bluetooth codecs

Which Bluetooth codec is best?

sdx15/Shutterstock

How to switch Bluetooth codecs

A woman identified as Jane Doe 4 has joined a lawsuit filed by three Tennessee teenagers against Elon Musk’s xAI over the role the company’s chatbot Grok allegedly played in creating child sexual abuse material.

According to a report in The Washington Post, the woman alleged that her stepfather used Grok to manipulate a photo taken when she was 11 years old to create more than 7,000 explicit images of her. The woman also said that her stepfather was found dead of suicide two days after the images were uncovered in a law enforcement raid.

“Limitless access to these tools is spreading so quickly,” said the woman. “It is taking everyday life and turning it into child sexual abuse.”

The teenagers who’d filed lawsuit accused xAI (now part of SpaceX) of failing to take basic precautions to prevent Grok from being used to create explicit images of real people, including minors. (X was flooded with millions of Grok-generated sexualized images earlier this year.) They are seeking class action status for their suit.

TechCrunch has reached out to xAI for comment.

If you are in a crisis or having thoughts of suicide, call or text 988 to reach the 988 Suicide and Crisis Lifeline.

DeepSeek has piqued the interest of the developer community by releasing an early version of its open source agent harness. This happens as harnesses have become increasingly important to those working with machine learning models.

“Powered by the Cordis meta-framework, DeepSeek Harness is an agent harness built around one core idea: Everything is a plugin,” the China-based AI biz said. “Models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are ALL implemented as plugins, and can be mixed, matched, replaced, and extended.”

The term “harness” came into common use this year to describe a longstanding software function – middleware or a mediation layer that handles the input passed to an AI model and the output returned from it. Harnesses oversee prompts, context management, tool orchestration, the agent loop, state management, error handling, safety, permissions, and related concerns.

Claude Code serves as a harness for Anthropic’s Claude model family and Codex performs a similar function for OpenAI’s GPT model family. And there are many other model harnesses, including Aider, Cline, Goose, OpenCode, OpenHands, and Pi, to name a few. 

The term isn’t precise: It may be used to refer just to the agent loop and tools, or it may be extended to a broader set of concerns related to orchestrating different tools, services, and capabilities like sandboxing, subagents, and so on. Google Antigravity, for example, consists of the Antigravity Agent Runtime (harness) that can be accessed through the Agent SDK, the Antigravity 2.0 desktop application, and the Antigravity CLI.

Vague definitions aside, AI model harnesses are now where much of the competition is happening, particularly as models proliferate and become commoditized. The harness often implements the user interface, a source of user inertia, and once developers configure their tooling and become accustomed to doing things a certain way, it becomes more burdensome to switch to a competing product, even if the interface consists mainly of a command line.

What’s more, various studies have suggested that model performance (and cost) varies significantly with the harness used, due to different design choices. For example, the Pi coding agent relies on a minimal system prompt of about 200 tokens. Claude Code by comparison uses a system prompt of around 10,000 tokens (or did until last month when Anthropic trimmed the system prompt by about 80 percent). The same model will produce different results with different harnesses.

DeepSeek Harness is noteworthy because of its innovative design, and because it shows Chinese AI labs moving to compete beyond model benchmarks and pricing.

First, it treats everything as a plugin. It uses the plugin system from its underlying Cordis framework, which is designed to make it possible to add and remove components dynamically without wreaking havoc.

“Plugins provide every agent capability, including models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI,” the DeepSeek Harness website explains. “Cordis services and events let the plugins work together. Developers can select, swap, or extend any capability in configuration without changing the DeepSeek Harness source code.”

A DeepSeek paper [PDF] by researchers Yifan Shi, Wei Zhang, and Tianyi Cui explains the function of Cordis in more detail. Cordis is designed to support dynamic composability – adding plugins and removing them on the fly without breaking the application.

The paper refers to this as temporal composability – removing a component and reverting its effect upon removal – and spatial composability – allowing components to manage dependencies upon other components.

It cites as an example the plugin system used by Microsoft’s Visual Studio Code. VS Code, the authors explain, runs all of its extensions in a shared process called the extension host. Once activated, they cannot be removed on the fly; the host has to be restarted. 

While VS Code provides a way for extensions to declare dependencies between extensions, it’s seldom used. DeepSeek Harness supports plugin dependencies.

The DeepSeek researchers argue temporal and spatial composability are necessary in a system where modification can occur continuously with little or no human oversight. It’s a way of avoiding forced restarts and crashes when components appear and disappear.

DeepSeek Harness also supports another useful feature: chain of thought traces.

“Everything the model sees is recorded in an append-only session log: system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection,” the DeepSeek Harness website says. “In the Trajectory view, you can inspect these records by source. Resume, fork, search, and replay all operate on the same event stream.”

DeepSeek R1 made waves when it was released last year and it was trained to use chain of thought reasoning. This involves breaking down prompts into a series of “thoughts” and reflecting on those steps before emitting a final answer.

Access to this intermediate reasoning turns out to be useful for assessing whether a model is reasoning well, whether its responses are accurate, how additional “thinking” affects output, and so on. 

Anthropic provides some access to thinking when extended or adaptive thinking is available (it varies by model). But increasingly the biz has been hiding model reasoning by summarizing chain of thought traces. That appears to be due in part to concerns that chain of thought traces can be used for copying models through a standard research process called distillation. 

Earlier this year, Anthropic said it had implemented classifiers for the “detection of chain-of-thought elicitation used to construct reasoning training data.” The company also does not display raw chain of thought. It explains that “the text in a thinking block is a summary of Claude’s reasoning.” Accessing raw thinking requires contacting Anthropic sales personnel.

Except for its open source models, OpenAI has also chosen to hide chain of thought reasoning, which the company uses for model monitoring. “After weighing multiple factors including user experience, competitive advantage, and the option to pursue the chain of thought monitoring, we have decided not to show the raw chains of thought to users,” the biz said two years ago when it introduced its o1 reasoning model.

With the newly released DeepSeek-V4-Pro and V4-Flash, the API provides thinking mode enabled by default. And as the open source model ecosystem matures, having access to chain of thought looks likely to become another opportunity for competitive differentiation.

“I don’t think the DeepSeek Harness is perfect but this is for sure the first time I have been looking at something new in the space and felt quite inspired to revisit some of our choices,” said Armin Ronacher, co-founder of AI biz Earendil, which now steers the development of the Pi agent, in a social media post. “I love that part about Open Source a lot!” ®

Anthropic wants to make AI-generated text easier to identify, and on paper, I have very little reason to complain. The company is experimenting with an invisible watermark that can be baked directly into text generated by Claude.

It sounds like a sensible idea. AI-generated text is everywhere, and knowing where something came from could certainly help. Moreover, Anthropic isn’t simply hiding a marker somewhere inside a document. Its approach changes how Claude selects words to create a statistical pattern that can later be detected.

But there is one detail that bothers me. Anthropic is testing just how persistent that watermark can be, even after the text has been modified.

That is where I can already smell trouble.

Claude touched my writing. Did it actually write it?

Think about translation for a moment. Let’s say someone writes an entire essay themselves in Spanish and asks Claude to translate it into English. The ideas are theirs. The research is theirs. The arguments are theirs. Claude’s only job is translation.

Yet the resulting text could still carry Claude’s watermark.

The same question applies to proofreading. What if someone writes something themselves and asks Claude to fix the grammar? What about shortening a paragraph, changing its tone, cleaning up dictated text, or simply making an awkward sentence easier to read?

These aren’t fringe uses for AI anymore. People increasingly turn to assistants like ChatGPT, Gemini, and Claude for everyday tasks that have little to do with generating original work. A watermark can tell you that Claude was involved with a piece of text. It cannot tell you whether Claude actually wrote it. Anthropic makes the same point, saying the watermark shows Claude’s involvement, not who created the original work.

Now imagine explaining that distinction to a professor after their detection software has just flagged your essay.

We already know how messy AI detection can get

I wouldn’t worry nearly as much if our track record with AI detection were particularly good. It isn’t.

MIT Sloan’s guidance is quite straightforward about existing AI detectors. It says they have high error rates and can lead instructors to falsely accuse students of misconduct.

We’ve already seen what that looks like in practice. Students have found themselves defending work they say they wrote themselves after automated systems identified it as AI-generated. In one case documented by The Guardian, a student’s essay was flagged as entirely AI-generated despite the student saying they had only used approved spelling and grammar assistance. The appeal was eventually accepted.

To be clear, Claude’s watermark is fundamentally different. Conventional AI detectors look at writing and essentially estimate whether an AI might have produced it. Anthropic is deliberately planting a detectable signal in Claude’s output. In theory, that should make its system considerably more reliable. But reliability isn’t the only problem here. Interpretation is.

We’re using AI to prove we didn’t use AI

Things have already reached a slightly ridiculous point.

Students worried about AI detection are turning to so-called AI humanizers, which rewrite text specifically to make it less likely to trigger detectors. Some students are even using these tools on work they wrote themselves because they’re worried about false positives. Detector companies, naturally, are developing ways to identify humanizers.

Read that again.

A human can write something, worry that an AI will think an AI wrote it, feed it through another AI to make it look more human, and then have yet another system determine whether the AI made it look human.

It’s a technological ouroboros.

Making Claude’s watermark resilient enough to survive editing and translation is technically impressive. Previous research has shown that translation can defeat some text-watermarking techniques, so solving that weakness would represent meaningful progress.

I just don’t think making the signal harder to remove solves the more important problem.

A watermark needs context

There are good reasons to watermark AI-generated content. It could help identify mass-produced misinformation, undisclosed synthetic text, or AI-written material that later ends up in training datasets.

The problem is that AI assistants now do far more than generate content from scratch. People use them to translate text, proofread documents, summarize research, help with code, improve accessibility, or simply clean up an email before sending it. In that context, detecting AI involvement does not automatically tell you who actually created the work.

All of those interactions involve AI to wildly different degrees. If Claude writes an essay from scratch, knowing that is useful. If Claude translates an essay someone spent three weeks researching and writing themselves, knowing Claude was involved tells you considerably less.

The watermark may be perfectly capable of answering “Did Claude touch this?” My concern is what happens when people start treating the answer as proof of “Did Claude write this?”

Anthropic can build the smartest watermark in the world. Unless the people using it understand that difference, I suspect we’re going to have some problems.