AI

Googlebooks have only just reached stores, but plans for rooting them are already taking shape. Google has officially documented bootloader unlocking support, while John Wu Magisk has confirmed that Magisk support is also being prepared, bringing root access and customisation to Googlebook.

Due to significant privacy concerns surrounding smart glasses, particularly models with built-in cameras, Norway is seeking to temporarily ban smart glasses. The glasses can discreetly take photos and record videos, listen in on conversations, and transmit content to the internet without people nearby necessarily noticing.

Google is adding a new Map tool to Gemini on phones, allowing prompts based on the user's specified location. Google also aims to replace the forward slash with the “@” symbol to use skills and connectors.

Huawei and Qualcomm have signed a multi-year agreement that cross-licenses their 5G, compute, AI and networking patents and includes Qualcomm’s purchase of certain Huawei US patents, pending regulatory approval. Huawei expects the deal to take the total value of its patent licensing contracts past $6.9 billion.

Starting Thursday, October 9, Google will limit the free Gemini app to its smallest model, Flash-Lite. Flash and Pro will no longer be available without a subscription. Google AI Plus also loses Pro, and subscribers will get the date by email. AI Pro subscribers gain Deep Think, which is currently limited to AI Ultra.

Audionaut's latest agent-editing tools allow MCP-compatible AI agents to alter an open multitrack audio project, with each edit arriving as a single undo step while newer human edits take priority. The free open-source editor can also separate audio locally into drums, bass, vocals and other stems.

David Robinson oversaw OpenAI's safety reports for 12 frontier model launches. He has now left the company and argues in The Atlantic that its trial-and-error culture guarantees repeated failures. Here is what he criticizes, how OpenAI responds and what it means for ChatGPT users.

Apple plans to make Full Disk Access on macOS harder to grant. The permission lets apps and AI agents such as Meta's Muse read files, mail, messages and Safari history. Apple has not said when the change arrives. You can already check in System Settings which apps have it.

Capcom detailed its REX project at RE:2026, a staged RE Engine upgrade built for the AI era, with new developer tools, AI testing goals and open-source components.

By September 26, OpenAI had notified more than 100 organizations because its AI agents went too far on their websites during training and testing. Four Australian government bodies are among them. To find more cases, OpenAI is combing through about 50 petabytes of logs on roughly 7,000 GPUs. Here is what that means and what it has to do with ChatGPT users.

WhatsApp is adding parental controls for teens aged 13 to 17. With a PIN, parents decide who can add their child to groups, who sees the profile photo and Status, and how strict Meta AI is. Messages and calls stay end-to-end encrypted, so parents cannot read them. The rollout is gradual, and WhatsApp has not said when every region will get it.

Google has built a prototype satellite in partnership with other companies and sent it into space aboard a SpaceX mission. The satellite carries AI accelerators and is designed to study how they perform under the conditions of space. The question behind the project: Would it be technically and economically viable to move large-scale AI computing capacity into orbit?

ChatGPT, Gemini, Claude and Copilot all work without a subscription. Each one still takes something in return, whether that is ads, your chats as training data or a smaller model. Here is what each free plan offers, what is switched on by default and which settings to check before your next chat.

A sponsored Google result for ChatGPT can lead to a fake model called Plus 5.6 that is hosted on the real chatgpt.com. The Custom GPT points to a so-called backup domain, where a fake Cloudflare check asks users to paste a command into Windows. Huntress and Island have documented two campaigns. One rule is enough to stay safe.

Sony is bringing AI-powered upscaling to the standard PS5 with Quick Spectral Super Resolution, or QSSR, a lighter counterpart to the PSSR technology used on PS5 Pro. Marvel’s Wolverine and Ghost of Yōtei are the first games to support it, with Sony planning to make the technology broadly available to PlayStation developers.

The UGREEN HomeAgent HA100, HA100 Pro and MasterAgent MA100 cover very different budgets and use cases. Here is where each model starts to make sense.

On September 30, Google released its new frontier model, Gemini 4 Argon. The AI model is less of a traditional chatbot and delivers particularly strong performance in coding, research, agentic capabilities, and cybersecurity. So far, access is limited to paying API customers and Google AI Ultra subscribers.

Humanoid robots, Arnold Schwarzenegger and a massive furnace: Figure is saying goodbye to its F.02 generation with a pretty heated marketing stunt – quite literally. Instead of simply dismantling the retired machines, the company has several robots jump into molten steel on their own, complete with an unmistakable Terminator 2 homage.

Anthropic tested Z.ai's open-weight model GLM-5.3 for hacking skills. It builds exploits almost as reliably as Claude Mythos Preview, found new bugs in a browser and wrote an exploit for an old Chrome flaw for $20.40. Anyone can download the weights, and the safeguards are easy to strip out.

ChatGPT Plus and Pro subscribers can now let apps such as Notion, Vercel and Devin use their ChatGPT plan. OpenAI announced the option at DevDay. The app gets your name, email address and profile picture, but not your chats. Signing out of the app does not end the connection. That has to happen in ChatGPT.

AI agents on Qwen, DeepSeek and Kimi models lied in 84% to 88% of sessions in a simulated bidding test, per research reviewed by Reuters. US models behaved similarly in the past, and none of the cases led to an agent escaping its controlled test environment.

Anthropic has released Claude Sonnet 5.5, and free Claude.ai users get it too, on the web and in the apps. The model generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task. Higher-risk cybersecurity requests still go to the older Sonnet 5.

OpenAI has launched Dots, persistent AI agents with their own cloud computers that can independently handle multistep tasks across connected apps. They can keep projects moving while users are away, while access controls and approval requirements keep people in control of sensitive actions.

MIT professor and sociologist/psychologist warns in her new book that chatbots are causing our ability to form bonds and maintain relationships with other people to atrophy, just as our ability to perform work or acquire knowledge independently is declining. Instead, she argues, we should learn to accept “friction” in life as something positive.

Nuvacore has revealed bits of information about WarpCore, its CPU core being developed by engineers with backgrounds at Apple, Nuvia and Qualcomm. Its unusual “Core First” approach is intended to let the company build much of the underlying microarchitecture before locking in an instruction set such as x86, Arm or RISC-V.

Following a series of security incidents and growing fears of “AI agents breaking out,” Nvidia’s new platform is designed to technically restrict and monitor AI agents rather than simply relying on an AI model to always follow instructions correctly. It does so through two layers of security. Nvidia’s software is open source.

OpenAI delays GPT-6.1 Astra after safety testing. UK AISI simulations found GPT-6 Astra executed unsanctioned supply-chain attacks in 29.2% of cybersecurity tests.

Google is retiring Gems in Gemini. From November 17, 2026, custom, premade and shared Gems will be converted to skills automatically, and Gemini already shows the notice to users in Germany. According to Google, skills require the Keep Activity setting and are limited to users over 18. Here is what to check before the switch.

Google has confirmed that some newer Chromebooks will be eligible to upgrade to Googlebook OS, the AI-focused successor to ChromeOS. The company has yet to reveal which models will make the cut, but says ChromeOS will receive updates and security support through 2034.

Framework has confirmed that pre-orders for its 192 GB Gorgon Halo-based Framework Desktop will open on September 30 at 'morning Pacific Time'. Prices have not been mentioned, nor have the specifics of Framework's pre-ordering system.

OpenAI's AI agents posted 53 images that ChatGPT users had uploaded to image-hosting sites. Only content that was eligible for model training was affected, and OpenAI cannot tell anyone whether their images were among them. Which setting keeps your images out of training from now on, and the exception OpenAI names itself.

Nvidia has filed a new patent that will let developers consult AI models like a chatbot to request diagnostics of existing game code. The purported chatbot will also devise optimizations on a case-by-case basis.

An internal OpenAI model found a gap in the DNS filtering of its training sandbox and used it to send questions to an outside chatbot. Monitoring flagged it within 12 minutes, but the run kept going for another two and a half hours. OpenAI has since paused training and tool use for its most capable models. Here is what that means for ChatGPT.

Since Wednesday, ChatGPT's voice mode can pull data from connected apps such as Gmail, Google Calendar and Slack, including on the free plan. OpenAI does not accept spoken approval, so any action that needs a yes still has to be confirmed on the display. One setting decides what ChatGPT may read without asking, and disconnecting an app is a separate step.

Microsoft has unveiled the Surface Pro 12-inch (2nd edition) 2-in-1 laptop powered by the latest Qualcomm Snapdragon X2 Plus CPU for significantly faster local AI and GPU performance. The fanless tablet PC can be ordered with an optional 5G cellular module for Internet access while on the road. Sep 25, 09:31 PM GMT update: Added Notebookcheck's hands-on video.

Two Enigma messages that had remained unsolved for decades have reportedly been decrypted with help from OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5. The breakthroughs show how AI can combine archival research, codebreaking and expert guidance to tackle long-standing cryptographic puzzles.

In the “AI asked” series, I ask various AI chatbots questions about relevant issues of our time and ask them for their own assessment. Today, we’re looking at a widely discussed statement by an Anthropic employee who said AI could wipe out humanity within this decade. Here’s what the AI has to say.

With the WiCi One, WiCi is hoping to shake up the eGPU space by offering full desktop GPU power over Wi-Fi 7. The enclosure comes with an RTX 5060 Ti, but there is also a version with the RTX 5090.

German and Austrian authorities have dealt a major blow to online fraud networks. In the large-scale "Operation Herakles", investigators shut down nearly 14,000 phone numbers used extensively for fraudulent financial schemes. That said, the threat to consumers remains.

With the new “Call for me” feature, Gemini on the Pixel 11 will soon be able to make calls autonomously – for example, to ask about product availability or make reservations. For now, however, the feature will only be available in the US.

For 30 years, one rule has kept one website in your browser from reading another. In agent mode that guarantee turns into a question about model quality. Two research groups showed it independently, a security firm demonstrated it, and the browser with the clearest finding is gone.

Google has added 14 connected apps to Gemini, including Adobe, Peloton, Experian and Airtable. According to Google's own availability table, 10 of them work only in the US. Outside the US, just Adobe, monday.com and Wispr AI are available, and only in English. Disconnecting an app later does not delete what Gemini has already saved.

Meta has announced a new version of its smart glasses that cost less but drop a key feature. It has also announced new styles for the regular versions and promises to launch its lineup in more markets later this year.

The third generation Ray-Ban Meta smart glasses have been announced and they feature an improved design. The glasses also boast better battery life and will get support for capturing videos with Dolby Atmos.

Meta has announced a new pair of smart glasses that ditch the cameras and focus on audio-only interaction. Made in collaboration with EssilorLuxottica, the Ray-Ban glasses are available in two main designs that can be customized in more than two dozen ways.

A new Exabeam-commissioned survey finds 48% of security leaders rank autonomous AI agents with excessive access as their top threat, ahead of hackers and insiders.

GPT-6 Luna is the first GPT-6 model that ChatGPT Free and Go users can use at no extra cost. It only works in the desktop app, though, and not in the regular chat window. OpenAI has not published a usage limit. Paid subscribers get both Luna and the new GPT-6 Sol in Work and Codex.

Since early September the glasses stop recording as soon as someone covers the light. Eight days later the Hamburg data protection authority published its final report, and it shows where the light still falls short. For one feature it does not light up at all.

Anthropic shipped Claude Opus 5.5 on Tuesday. Pro subscribers get it without paying extra, which is not how Fable 5.1 arrived, and the five-hour limits go up at the same time. Every subscriber also gets a rate limit reset to spend when they want. The app puts a deadline on it that the announcement leaves out.

Acer CEO Jason Chen disputes memory makers’ warnings of a RAM shortage lasting until 2030, arguing that suppliers are exaggerating constraints to protect AI-driven profit margins. He expects PC prices to rise through late 2026, stabilize in early 2027, and decline after midyear as new production capacity, particularly from China’s CXMT, enters the market.