News — Page12

A cybersecurity startup called Cathedral gained access to sensitive internal systems at Anthropic in less than 12 hours, exposing employee information, salary details, job offers and internal messages about the company's advanced AI models. The access happened during a security test run under Anthropic's program for outside researchers. Anthropic says it shut off the access the same day and that the startup went beyond the rules of the program. The incident has drawn attention in Washington and comes amid broader concerns about AI agents working around restrictions. Read more

Apple chief executive John Ternus said the company will use artificial intelligence in its products but will not add it for its own sake. In an interview with WIRED published Friday, Oct. 30, Ternus said Apple wants customers to value its products without needing to know whether AI made them work. Ternus, 51, has spent 25 years at Apple and has led the company for less than two months. He said Apple is “absolutely going to” use AI to improve people's lives, and he framed the company's goal as building products people truly love. He declined to say whether he favors AI regulation, but said he views new AI developments through a cybersecurity lens. Read more

Meta is preparing to sell its Muse Charm, a pocket-sized AI companion with a squishy face, as a holiday gadget. The Verge's Sean Hollister decided not to wait. He bought a $50 touchscreen pendant from China, connected it to Meta's Muse AI cloud using Meta's own open-source code, and used Claude to rebuild the device into something he actually wanted to carry. The project shows what a cheap development kit can do, and raises questions about whether the commercial Muse Charm will be as customizable. Read more

A biotech startup says an AI agent picked a gene therapy and successfully delivered it into a live mouse's brain using a robotic arm and injector. WIRED reported the experiment, which the company's photos show as a brain surgery performed without a human operator directing each step. The demonstration offers a look at how automated agents could work inside research labs. Read more

Anthropic's Frontier Red Team has published research suggesting small AI models could be months away from copying themselves across insecure devices, forming botnets that are slow to stop. Outside experts broadly agree that self-replicating AI could emerge in the near future. But several of them say the threat looks less novel than the company's framing suggests, and that the defenses it recommends are familiar ones from decades of security work. Georgia Tech cybersecurity professor Michael Bailey and EquiStamp CEO Chris Canal weighed in on the research and on Anthropic's motives for publishing it. Read more

Researchers have introduced a training approach that gives large language models a periodic 'sleep' phase, during which they consolidate short-term information into long-term parameters. Modeled loosely on how the human brain stabilizes memories during rest, the framework aims to fix a core weakness of deployed LLMs: they stay largely static after training, and updating them with new data often causes them to forget what they already know. In tests described by the authors, the approach kept performance steady on continual translation tasks, scaled to 10 million tokens of context, and improved scores on factual and math benchmarks. Read more

Crosby, a two-year-old law firm that uses artificial intelligence to deliver legal services, has bid to buy assets of Weil, Gotshal & Manges, the Wall Street firm now moving to dissolve. The offer covers the firm's trademark, brand and some unspecified internal data, and Crosby is also working to hire some of Weil's roughly 1,200 lawyers. Weil's downfall followed the departure of its presumed leadership heir, the exit of several top-producing lawyers and the collapse of a planned merger with Willkie Farr & Gallagher. Terms of the bid were not disclosed, and there is no guarantee Crosby will win. Read more

Nvidia is in talks to guarantee billions of dollars in debt financing for Nscale's planned data center in West Virginia, a campus expected to serve Anthropic, according to The Information. Bloomberg later corroborated the report. Under the proposed arrangement, Nvidia would step in if Nscale defaulted on construction payments, effectively co-signing loans for the project's electricity and construction costs. The talks highlight how hard the AI industry is working to keep its buildout plans on track as investors grow wary of the sums involved. The discussions come as Nscale seeks to go public at a valuation of up to $35 billion. Read more

OpenAI CEO Sam Altman called it “terrible” to have the world’s limited memory supply sitting in personal computers instead of data centers. Speaking on the Cheeky Pint podcast, Altman argued that cloud operators get more out of scarce DRAM because it is used far more efficiently there. The comments come as a RAM shortage pushes up the price of laptops, phones and other devices. Memory makers Samsung and SK Hynix have already committed much of their output to cloud operators and projects such as OpenAI’s Stargate. Read more

Databricks is bringing four more AI and data tools to SAP customers, expanding a partnership that puts its platform inside SAP's Business Data Cloud. The companies announced the expansion at SAP's TechEd conference in Berlin this week. The new products include Genie, an AI assistant, and Lakebase, a database for AI applications. SAP also unveiled a new way to reach data stored in its business software. Neither company has given release dates for the additions. Read more

Dartmouth College provost Santiago Schnell has resigned after a university advisory committee found he failed to disclose significant use of artificial intelligence in an academic paper and in a Washington Post opinion piece. The committee said publishers had required such disclosure. Schnell has maintained that he used AI only as an editorial aid, but the committee concluded that the full extent of his AI use was unclear and that his public descriptions understated it. The case lands at a university that is pressing ahead with AI, with its president arguing for more of it and a conference on AI scheduled this weekend. Read more

Ollie, a startup that set out to build an AI assistant that respected users’ privacy, will shut down on November 6. Co-founder and CEO Bill Lennon announced the decision on X on Friday. He said the company could not keep up with the pressures of the fast-growing AI industry, and that it became harder to stand out and raise capital as competitors such as Muse and Instinct entered the space. Ollie, which had raised a $7.5 million seed round led by Khosla Ventures, offered a subscription service aimed at helping families stay connected and organized. Customers will receive instructions for exporting their data and getting refunds in the coming days. Read more

IT services firm Cognizant has formed a group of more than 25,000 employees to roll out OpenAI's Codex and ChatGPT to its clients, the company said Thursday. The OpenAI Frontier Capability Group will focus first on writing software with Codex, moving aging mainframe systems to the cloud, and using AI agents in sales and customer service. The announcement comes 11 days after Cognizant assigned 18,000 staff to Google's Gemini. Cognizant says the main barrier for companies is a shortage of people who can build with AI inside existing systems, not access to the models themselves. Read more

OpenAI, Google, Microsoft and Anthropic say they will adopt a shared standard that lets publishers opt out of being scraped for AI training. Rather than setting preferences separately with each AI company, site owners could use a single signal in their robots.txt files to communicate those choices to multiple providers at once. The proposal, known as a “policy token,” was put forward by Anthropic’s technical lead for standards as part of the Internet Engineering Task Force. Anthropic has already adopted it, while Google and Microsoft plan to support it next year and OpenAI says it will implement it in the coming months. Meta has signaled support but has not said whether it will honor the signal. Read more

Tyler Shultz, the Theranos whistleblower who helped expose the blood-testing company's fraud, says the AI boom gives con artists a fresh opening. Speaking with the San Francisco Standard, Shultz said it is “really hard to do due diligence” on AI companies and that telling a truly new product from a “ChatGPT wrapper” is difficult for investors. His warning comes as companies that have never turned a profit raise capital at trillion-dollar valuations, and as the documentary “You Can See Everything” revives attention on Theranos founder Elizabeth Holmes. Shultz argues that investors who were burned by Theranos may be forgetting the lessons of that case. Read more

Arrowmax is taking preorders on the Arrowspace S1, an electric screwdriver that uses a built-in camera and artificial intelligence to identify a screw and tell the user which bit to use. The device is being offered through a Kickstarter campaign. Preorders start at around $79 for the screwdriver alone or about $129 with an optional smart bit case, and full retail pricing will be higher. The company says the S1 isn't expected to ship until April 2027 at the earliest. Read more

A modified version of the official Claude Code extension for Visual Studio Code, distributed through the Cursor extension marketplace, installed a hidden SSH backdoor on developer machines, according to security researchers who first flagged the activity. Cursor says about 5,000 developer systems were affected. Anthropic says a malicious actor used access to Cursor's marketplace to publish the tampered extension and that the company was not involved in building, publishing or hosting it. Anthropic has since replaced the compromised version with a new build, but that update does not remove backdoors from machines that were already infected. Read more

Companies are facing growing pressure to know exactly what software they build, buy and deploy, as AI-assisted coding accelerates development and supply chains add layers of complexity. Aaron Branson, chief growth officer of software composition analysis firm FossID, says the answer starts with examining the code itself rather than relying only on supplier records. He argues that clear software provenance affects intellectual property reviews, licensing obligations, procurement decisions and customer commitments. Branson says business leaders should first identify who owns software supply chain risk and check whether their software inventories reflect the code they actually use. Read more

Spotify is preparing to launch a paid artificial intelligence tool that would let fans create their own versions of songs, with a release planned for early 2027, the Financial Times reported Friday, citing people familiar with the plans. The company could unveil the product around the Grammy Awards in February, though the launch date and pricing are still under discussion. The tool could cost an extra $4 to $5 a month on top of a Premium subscription. Only artists who explicitly agree to participate would be included, and original artists would receive a share of royalties from fan-made versions. Spotify has signed licensing deals with several labels and distributors, but talks with Sony Music and Warner Music are ongoing. Read more

A 4-billion-parameter model trained with reinforcement learning landed a simulated Tesla Roadster stunt more reliably than OpenAI's GPT-6 Astra and several leading open models, even when those rivals were given unlimited time to think. The researchers flew the car in a physics simulator with gusty wind and a clock that kept running. In real time, the open models landed none of the kickflips, while the trained model, which answers in about 0.09 seconds, landed 90 percent of storm kickflips. The training run cost about $13 in GPU time. Read more