The gap between Mac and PC is finally closing, but be careful before you pick a Mac as your primary gaming platform.

What a MacBook can and can’t handle

Filipe Espósito for Engadget

Which MacBook to buy, and when to look elsewhere

Filipe Espósito for Engadget

Security teams at Google, JPMorgan Chase, Weaviate, France’s interministerial digital directorate (DINUM) and the city government of Tangerang in Indonesia have each fixed the same type of flaw in their Model Context Protocol (MCP) servers. Independent researcher Syed Anas Mohiuddin reported all five, he wrote in an update published this month.

MCP is the standard AI agents use to call tools and data sources. The flaw is server-side request forgery (SSRF). An MCP server takes a URL, path or endpoint from an agent and builds an outbound request from it, without checking where the address actually resolves. That lets whoever steers the agent decide what the server talks to, including internal systems.

In May, Mohiuddin argued that the problem was structural. If it was, he predicted, teams that share no code or owner would all produce it.

“Watching the same mistake come back from a hyperscaler, a bank, and a national government, one report at a time, is the moment the May argument stopped being a guess,” Mohiuddin wrote.

Five fixes

Google’s MCP Toolbox for Databases had an HTTP client with no restrictive redirect policy and no check on target IP addresses, according to the GitHub advisory. A crafted path parameter could send its requests to internal or external endpoints. The flaw, CVE-2026-14540, carries a high rating of 8.0 and affects versions 0.3.0 to 1.4.0. Google’s fix adds a guard against DNS rebinding and lists of allowed and blocked IP ranges, and credits Mohiuddin.

JPMorgan’s open-source repository includes a documentation-search MCP server with two tools that fetch content. One checked domains against an allowlist. Its sibling fetched any URL the caller supplied, Mohiuddin wrote. JPMorgan forked the component from an AWS project that never fetched the caller’s URL at all. JPMorgan’s Responsible Disclosure team confirmed the finding and deployed a fix, he wrote. The finding is medium severity, he wrote.

Weaviate restricted its Google module’s endpoint settings to Google API hosts, he wrote. DINUM’s official MCP server for France’s open-data platform fetched URLs supplied by data producers, which could point at internal or cloud metadata addresses. Its fix, titled “harden SSRF on external APIs”, opens with “Reported by Syed Anas Mohiuddin”.

In Tangerang’s Wazuh MCP server, a tool advertised SSRF protection but only rejected literal IP addresses, according to a high-severity advisory published on 3 September. It never resolved hostnames.

Rapid7 fixed a different bug he found, CVE-2026-97228, in its Bulk Export MCP server. The bug allowed GraphQL injection within the operator’s own access, and Rapid7 rates it low, at 2.7, its database entry says.

Still open

On 2 September, Mohiuddin privately reported issues in five MCP servers under the US General Services Administration’s Technology Transformation Services. They include servers for Veterans Affairs benefits claims, CMS Blue Button, regulations.gov, USASpending and CDC PLACES. All five are still in triage and not fixed, he wrote.

In the Veterans Affairs case, the server logs full error responses from the benefits API without redaction, according to Mohiuddin. Those can contain a veteran’s name, Social Security number, date of birth and address. He is withholding code-level detail until maintainers patch the servers. His report on Japan’s Digital Agency grants server, which had no authentication, also remains open.

‘Protocol pivoting’

Mohiuddin calls the wider attack class “protocol pivoting”. An attacker plants text in content an MCP tool returns, shaped like a task for Google’s A2A protocol. An orchestrating agent passes it to a subagent, which runs it because it trusts the orchestrator.

“Every piece in that chain did exactly what it was designed to do, which is what makes this so tricky to catch,” Douglas McKee, Rapid7’s director of vulnerability intelligence, told Ars Technica.

Markus Vervier of X41 D-Sec told Ars that the technique is a form of indirect prompt injection. Mohiuddin will present the findings at MCPCon North America in San Jose on 23 October.

This year, the United States hit a milestone nobody asked for: more measles cases than any year since 1991, six decades after a vaccine was first developed. Fear and misinformation convinced families to walk away from settled science. Children paid the price.

I bring this up not to talk about vaccines, but to talk about literacy.

On July 15, Lynn Vasquez of the New Mexico Public Education Department met with lawmakers alongside Mark Angel, CEO of Amira Learning, to defend the use of the company’s AI-powered literacy screener in public schools across the state. New Mexico has been under a court order since 2018 to improve outcomes for its most underserved students, and officials turned to Amira in 2025 to help improve reading proficiency.

The software lets teachers “listen” to 5- and 6-year-olds as they read aloud, catching the hesitations, substitutions and decoding errors that reveal struggle but are easy to miss. It allows educators to intervene early and more accurately, delivering on the promise of the science of reading. The state initially required districts to administer the Amira Reading Assessment to K-2 students at the beginning, middle and end of the school year.

Meanwhile, parents in cities including Santa Fe and Albuquerque have organized in opposition to the tool, urging their districts to opt out of participation – an option now available across the state. Their objection isn’t with reading assessments, but with data privacy fears and a growing discomfort with anything called AI.

In September, that argument stopped being local. New York City, the largest school system in the country, announced a one-year moratorium on student-facing generative AI for roughly 600,000 students in 2-K through eighth grade. Days later, Los Angeles pulled generative AI from every student’s district device. Both policies target generative tools like chatbots. Amira isn’t one of them: it listens to a child read aloud and flags where they stumble. But once “AI” becomes the enemy, a reading screener lands in the same bucket as ChatGPT. 

Whatever you think of these policies, the reasoning coming out of New York and Los Angeles will be quoted in state capitals for the next year, New Mexico’s included.

The same machinery turns in both stories: settled science on one side, fear and misinformation on the other, children holding the bill. I have seen this movie. I know how it ends, and who pays.

I spent 15 years as a Google executive championing classroom technology, and I’ve been called a techno-optimist more than once. Most education technology has not earned your trust: schools have been sold engagement dressed up as learning, apps that babysit rather than teach.

Classroom technology demands skepticism that sorts products without evidence from the rare ones with it. Yet when skepticism stops sorting and starts banning, it becomes panic. Fear in education carries real costs. In New Mexico, it is tallied in third graders who can’t read.

Politics have derailed settled reading science before. In 2000, the National Reading Panel reviewed the research and found that children learn to read best with explicit, systematic instruction in phonics and phonemic awareness. Washington funded that science through Reading First, then killed the program after the Education Department’s inspector general found officials had steered schools toward favored products. Emily Hanford’s 2022 podcast, “Sold a Story,” forced a reckoning. As of November 2024, more than 40 states have passed science of reading laws.

A Costly Experiment

New Mexico cannot afford to run the same experiment. The state’s fourth graders post the lowest reading scores in the nation. According to a study commissioned by the Annie E. Casey Foundation, children not reading on grade level by the end of third grade are four times more likely to drop out of high school. In New Mexico, that’s not a statistic. It’s a pipeline, and everybody knows which kids are in it.

The state’s own numbers show the technology is working. In July, the NMPED shared a report with lawmakers that revealed the impact of Amira during the previous school year. The screener flagged 16.7% of first graders, roughly one in six, as showing signs of dyslexia, in line with national estimates. Under the previous computer-based screener, the figure reached 45%. Nearly half the state’s six-year-olds carried a label they didn’t deserve, often because their first language wasn’t English, and the old test couldn’t tell the difference.

Fewer false positives means intervention dollars reach children who actually need them: a machine enabling judgment, not replacing it.

Of course, critics raise legitimate questions. Thirty-five New Mexico lawmakers sent a letter to the state’s education department wanting to know what happens to children’s voice recordings. Los Alamos and Santa Fe paused the screener over parent concerns that Amira was training its products with the recordings for commercial use. 

Any parent can fairly ask about data privacy or whether the tool struggles with kids who have accents or speech differences – and how that’s being addressed. AI speech-recognition tools sometimes fail to accurately assess speakers who have accent or speech differences because they were not specifically trained to recognize different accents. Those challenges deserve vendor testing and public answers, not a shrug.

But I was that kid. I grew up in Hell’s Kitchen, raised on welfare by a single mother from Argentina. English was something I assembled at school, piece by piece. The analog status quo sorted kids like me quietly, by teacher impression, by gut feel: no audit trail, no accuracy study and no appeals process.

Human judgment has biases, too, but you can measure a machine’s error rate and fix it. You cannot audit a hunch. If the tool mislabels children with accents, prove it and demand the fix. That’s governance; pulling the tool just hides the bias where nobody can see it.

Uneven Stakes

The stakes are not evenly distributed. Some families organizing opt-outs may have the books, tutors and confidence to work the system. The children that New Mexico is failing have a screener, and may have nothing else. A kindergartner cannot opt herself back in at 16, when the problem nobody caught becomes a life problem everybody can see.

By all means, govern. Write the rules of procurement. Tell parents in plain language what gets recorded, how long it’s kept and who can use it. Demand the data of the software’s accuracy and throw out every product that can’t produce it.

Albuquerque shows what that could look like. This school year, the district kept the screener, communicated with parents about the rationale for voice recordings and opted to store recordings for a shorter time period.

Every state is now going to be asked whether it wants what New York and Los Angeles want. New Mexico’s problem is not AI. It’s that most of its children aren’t reading proficiently. Solve that one first. 

Jaime Casap spent 15 years as Chief Education Evangelist at Google, advising school systems and education leaders on classroom innovation and technology policy. Raised by a single mother in Hell’s Kitchen, he is the co-author of “Our First Talk About Poverty.”

On Friday morning, Shivon Zilis, a top executive and lieutenant of Elon Musk at his brain chip implant company Neuralink—as well as mother of four of the 14 children he is known to have fathered—took to his social media platform, X, to announce an abrupt shift in their relationship.

“It’s hard to go from in love to let go in a week with no warning, but that’s just how it is sometimes,” Zilis wrote, sharing a screenshot of the X account @BigTechAlert that showed Musk had unfollowed her there. Zilis’ post continued: “I feel happy that I got to show Elon how it feels to be deeply understood, nurtured, and peacefully loved with stability and gentleness,” followed by warm praise and best wishes for Musk. “I do hope and pray he and I can be great friends and coparents,” Zilis concluded, leaving some ambiguity as to what exactly changed between them and whether she remains a director of operations and special projects at Neuralink, though it does not appear as if she has left the position.

While this was a strange revelation of a private matter, it seemed hardly out of the ordinary for Musk’s dizzyingly complicated family life. At this point, all four of the women confirmed to have had children with him—Zilis, ex-wife Justine Wilson, Claire Boucher (better known as Grimes), and Ashley St. Clair—have used X to share details of their time with him or push back against his own claims.

Both Wilson and St. Clair participated in filmmaker Alex Gibney’s sprawling new documentary Musk, in theaters nationwide on October 16 and screening at the WIRED World Fair in Miami next month. St. Clair, a former conservative influencer, alleges that she rejected a nondisclosure agreement worth $40 million that would have prevented her from talking about her tumultuous association with Musk. Among an ensemble of tech entrepreneurs, journalists, and former Musk employees, the pair stand out as arguably the most important critical voices in the film, shedding light on Musk’s rise in Silicon Valley as well as his turn toward MAGA extremism during the 2024 election.

At a run time of nearly four hours, Musk sounds like daunting viewing, though once it’s over, you’ll find yourself wondering if it could’ve been a 10-part miniseries. That’s how much ground there is to cover, from the unrealistic ambitions of landing on Mars and the corporate colonization of Texas to the fatal collisions involving “self-driving” Teslas and the annihilation of the US Agency for International Development by the so-called Department of Government Efficiency (DOGE) at the beginning of Trump’s second term. (WIRED’s reporting is cited in the documentary, which features interviews with global editorial director Katie Drummond and contributing editor Zoë Schiffer.)

Alongside a fast-moving, comprehensive account of Musk’s career as an all-powerful and untouchable oligarch, his former romantic partners provide crucial context about exactly who he is, piecing together a devastating portrait of the type of man to “let go” of a significant other by unfollowing her on social media.

Your phone can fold in half, but your keys shouldn’t join it.

Treat the inner screen with care

What is the lifespan of a foldable phone?

MyImages – Micha/Shutterstock

In the world of home theater, there are traditional A/V systems and powered speakers, and there are soundbars. Then there’s the Nakamichi Dragon, a premium-priced soundbar system designed to bridge the gap between those worlds like no other setup I’ve encountered.

I reviewed the latest version, which includes a single 12-inch subwoofer in an 11.1.6-channel configuration and lists for a whopping $3,800, although it often sells for considerably less. The Dragon delivers exhilarating performance, especially with 3D audio formats like Dolby Atmos, but before you consider buying one, you’ll want the whole story.

Skip the reading? Click play above or watch the Nakamichi Dragon review on YouTube.

The Dragon’s sheer size and weight make it a real chore to unbox and set up. The soundbar alone weighs more than 32 pounds, and at 58 x 4.4 x 7.7 inches (W x H x D), it barely fit on my mid-size console. The subwoofer weighs about 80 pounds and extends nearly 18 inches deep.

The surround speakers are much easier to manage, although they feel a little cheap considering the price. They do offer some interesting technology, including a top-mounted dial that lets you physically adjust the direction of the upfiring drivers. The overall angular design is big and bold and practically demands an equally large TV to match.

The 11.1.6-channel model I tested comprises 28 speakers, including the 12-inch subwoofer, and you can add up to foursubs if you dare. The main soundbar houses 17 drivers, including four upfiring drivers for height effects and three 1.5-inch AMT tweeters. Each surround speaker adds two more 1.5-inch AMT drivers, two 3.5-inch drivers for side and rear surround effects, and a 3-inch upfiring driver.

Nakamichi rates the system at a claimed 3,000 watts of total power. Connections include three HDMI 2.1 inputs, one HDMI eARC port, optical input and output, 3.5mm analog input and output, and Bluetooth 5.0 with aptX HD. Oddly, there’s no Wi-Fi for features like Spotify Connect or AirPlay.

Audio support is extensive, including Dolby Atmos via Dolby TrueHD and the major DTS formats, including DTS Pro.

Controls and setup are handled through a loaded, old-school remote and an on-screen display that looks like it was pulled straight from a legacy A/V receiver. There’s no companion app for deeper adjustments, another odd omission at this price, and no automatic room calibration.

Instead, you can select from preset room sizes or manually enter your room measurements, which I recommend for the best results. You’ll still need to tweak several settings, especially the subwoofer, which is far too loud by default in a medium-sized room.

Those adjustments live inside the Auto Setup and Calibration menus. The available settings are extensive, but also fairly complicated, so expect to spend some time getting everything dialed in.

Performance is full throttle, from clear, forward dialogue to high-flying — and impressively precise — overhead effects and bass that can literally shake your windows. The low-end punch was far too much for my room before adjustment, although the system is admittedly better suited to larger spaces.

The overhead effects can be similarly aggressive. You can always dial them back, but it is undeniably fun to hear emphasized details like the almost visceral hull creak during the crash scene in Guardians of the Galaxy Vol. 2.

My biggest criticism is the overall sound signature, which can feel too thin and tightly wound, especially with lower-quality content. It lacks the full-bodied warmth of traditional speakers and even some less expensive soundbars, including Samsung’s HW-Q990.

Music is also less convincing than I expected at this price. The Dragon can be spectacular with movies, but it is not the system I would choose if two-channel music listening were a major priority.

The Bottom Line

If you’re after unearthly bass, cinematic surround sound, and powerful, almost surreal height effects, the Dragon is worth considering. I still take issue with its overall finesse and warmth, especially through the midrange, but there’s no denying how detailed, dynamic, and immersive it can sound with 3D audio soundtracks.

I think many A/V enthusiasts will find a better return by going with moderately priced speakers from brands like Klipsch, Polk, or SVS and a traditional A/V receiver. On the other hand, folks who want simplicity will find it in more versatile setups like Sonos’ Arc Ultra ($1,099) with a Sonos sub ($899) and Era 300 speakers ($499 each), or more affordable packages like Samsung’s HW-Q990.

That said, you’ll have a hard time finding this level of intensity, immersion, and sheer bass impact from another all-in-one home theater system. If that kind of larger-than-life presentation is what you’re after, the Nakamichi Dragon delivers in a very big way.

Pros:

• Massive, earth-shaking bass
• Powerful and adjustable overhead effects
• Clear dialogue even in chaotic scenes
• Loads of physical inputs
• Deep-dive settings and loaded remote

Cons:

• Difficult physical setup and complicated settings
• Sound signature is thin and forward
• No Wi-Fi or auto-calibration
• Cost is similar to traditional A/V setups

Our Ratings:

★★★★★★★★★★ Performance

★★★★★★★★★★ Build Quality

★★★★★★★★★★ Features

★★★★★★★★★★ Usability

★★★★★★★★★★ Value

Where to Buy:

Related Reading:

OpenAI is preparing to add invisible watermarks to text generated by ChatGPT and Codex in the European Union.

The watermark will not be visible when you read or copy the text. Instead, OpenAI says its new textGrain technology slightly changes the model’s word choices to create a statistical pattern that can later be detected.

“Over the coming weeks, we will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union,” OpenAI explained.

OpenAI is not making this a global default yet.

Starting today, API developers worldwide can opt in to watermarking for supported models, but it remains disabled by default.

The company is also opening applications for its watermark detector, although access will initially be limited to approved researchers and expert organizations.

OpenAI admits its AI watermark can disappear when you edit the text

Text watermarking is far from perfect, and OpenAI’s own tests show that fairly normal editing can significantly reduce its ability to detect AI-generated text.

“In an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words reduced it to 17%,” OpenAI noted.

At a 1% false-positive target, OpenAI detected the watermark in about 80% of 200-token psychology responses, compared with roughly 95% when the text reached 400 tokens.

Detection was even worse for subjects such as mathematics, where the model has less freedom to choose different words.

“The absence of a detected watermark does not prove human authorship,” OpenAI warned. “Text generated with OpenAI tools may be too short, edited, or translated for detection to work reliably.”

OpenAI also says a detected watermark does not reveal who generated the text, their account, prompt, or conversation, and it cannot tell how much of the final work was written or edited by a human.

Interestingly, OpenAI says watermarking does not meaningfully affect the quality of GPT-6 Astra, with benchmark results remaining broadly similar when textGrain is enabled.

Join Mikko Hyppönen and security leaders from the NFL, CHANEL, and Atlassian for a two-hour digital summit on what AI-speed attacks change, what defenders should stop doing, and how to validate, decide, fix, and re-validate at machine speed.

Save your seat