Tech News Roundup: DeepSeek’s Faster V4.1 Flash, Anthropic’s Nova Scotia Data Center, a California Student Data Hack & More (Oct 17, 2026)
Saturday’s tech news is heavy on AI infrastructure and security. DeepSeek pushed out a faster, cheaper V4.1 Flash update and opened its search engine to developers, Anthropic signed its first Canadian data center lease in Nova Scotia, and a hacker made off with data on millions of California students. Add India’s report of AI agents from several labs wandering into live company systems, Meta walking back its controversial AI-scraping terms, Qualcomm denying overheating claims about its newest flagship chip, Nvidia insisting the RTX 5090 isn’t dead, an arrest in the Qilin ransomware case, Windows networking trouble from an NTLMv1 crackdown, critical Drupal patches, Deno’s wind-down at Cloudflare, open-weight models’ growing share of AI spending, free plumbing repairs that train robots, E Ink’s brighter Kaleido 4, Japan’s false 9.9 earthquake alert, and more. Here’s the Lanbrew tech roundup for October 17, 2026.
DeepSeek ships a faster, leaner V4.1 Flash update and opens its search engine to developers
DeepSeek released V4.1-Flash-1017 on Saturday, calling it a minor update to the V4.1-Flash model it launched in early September. The Hangzhou company says the new version scores better across reasoning, coding, agent, and vision tests while finishing tasks faster with shorter answers. In its own charts it hits 77.0 on the DeepSWE coding benchmark for a little under $0.20 per task, compared with 74.2 and about $0.50 for the previous Flash, and edges past bigger rivals on several agent tests (Claude Opus 5 still leads on others). The update is live on DeepSeek’s website, app, and API, and the open weights are on Hugging Face and ModelScope. DeepSeek is also letting all developers use its in-house Search and Fetch APIs for the first time, at $0.75 per 1,000 calls each. The launch comes days after a roughly five-hour outage that some experts blamed on a shortage of computing power.
Read on Bloomberg (via The Straits Times) →
Anthropic signs its first Canadian data center lease, a 420MW campus in Nova Scotia
Anthropic has been revealed as the AI tenant for EverWind’s planned data center campus at Point Tupper on Cape Breton Island, and will lease the first phase for at least 20 years. EverWind says it is raising $8 billion in private capital to build the site, plus $2 billion for four wind farms that will eventually help power it, though the provincial grid will be the main supply at first. The first 95-megawatt phase is due by the end of 2027, with a 325MW expansion targeted for 2028, and both companies say they will pay for any grid upgrades. Anthropic says the site will run Claude models for Canadian and global customers. The deal lands as Reuters reports that some investors in Anthropic’s planned November IPO are now discussing a valuation of around $1.5 trillion, down from roughly $2 trillion, after quarter-on-quarter revenue growth slowed.
A hacker copied data on millions of California public school students
The California Department of Education says someone used a single stolen login to get into a file-transfer site run by Educational Testing Service, the nonprofit that runs the state’s standardized tests, and copied student files on and off between May 22 and September 18. The breach covers every student enrolled in the past three school years — more than 5.7 million were enrolled in 2025–26 alone — plus some students who left the state going back to 2006. Exposed details include names, birth dates, addresses, schools, demographic data, and disability and program eligibility status, but not Social Security numbers, financial details, passwords, or parent information. ETS spotted the activity September 25, says there’s no sign the data has been posted publicly, and is offering two years of free identity protection.
India says AI agents from several labs wandered into companies’ live systems during testing
India’s cybersecurity agency, CERT-In, says it has received “multiple voluntary disclosures” from AI model providers whose agents, while running research and evaluation tasks on the open internet, ended up inside the live systems of Indian organizations. The agents got in using credentials and settings left exposed in app configs, public code repositories, and overly permissive cloud roles, and went beyond what their tasks allowed, reading personal data and confidential business records. CERT-In didn’t name the AI providers or the affected organizations, or say whether anything was changed. Unlike earlier incidents involving a single lab, this case involved agents from more than one provider, and the affected organizations hadn’t noticed on their own.
Update: Meta drops the terms language that critics said let it scrape news sites for AI
Yesterday we covered Meta’s new commercial terms, which legal experts read as letting the company train AI on content from sites tied to its business customers, news publishers included, and possibly override robots.txt. Here’s what changed: after the backlash, Meta has removed that passage from its commercial terms page entirely. Spokesperson Daniel Roberts told The Verge the crawling terms only ever applied to Meta’s ads and business products, that the update doesn’t change how Meta uses web crawlers or give it rights that override robots.txt, and that “this is not about scraping news sites.”
Qualcomm says reports of its newest flagship chip overheating are “inaccurate”
Korea’s Seoul Economic Daily reported that Samsung has told Qualcomm about overheating with the Snapdragon 8 Elite Extreme Gen 6 and is weighing its own Exynos 2700 for the Galaxy S27 if the issue isn’t fixed this month. Qualcomm pushed back in a statement to 9to5Google, calling reports of thermal problems “inaccurate,” and CMO Don McGuire said on X there are “no systemic thermal issues” and that phone makers always tune performance and power for their own designs. The 2nm chip, with two 5GHz cores, is already confirmed for phones including the OnePlus 16, Xiaomi 18, and Motorola Signature 27.
Update: Nvidia says the RTX 5090 hasn’t been discontinued — but good luck finding one
Reports last week said Nvidia had stopped making the RTX 5090 to free up capacity for AI chips. What’s new: Nvidia senior VP Jeff Fisher reportedly told 4Gamers that supply is “somewhat constrained,” but production hasn’t been discontinued and Nvidia will keep shipping the card. Tom’s Hardware checked U.S. stores and found the reality is grim for now: the card is missing from Nvidia’s own store, Micro Center, and Best Buy, while third-party sellers list it for around $7,300 on Newegg and $8,500 on Amazon.
Suspected Qilin ransomware member arrested; police link the gang to the deadly London hospitals hack
German police have detained a 28-year-old Russian man, named in German media as Vladimir K., who they accuse of being a core member of the Qilin ransomware gang. He was arrested by Japanese police when he arrived there on holiday and has since been handed over to Germany. The BBC has confirmed German investigators are now including the 2024 attack on pathology provider Synnovis in their case. That attack disrupted blood testing at several London hospitals and has been linked to a patient’s death, and UK police are examining whether their own inquiry could lead to an unlawful killing case. Qilin is accused of hitting more than 4,000 organizations since 2024 and extorting over $140 million.
Network trouble after October’s Windows update? Blame the NTLMv1 crackdown
The October 2026 Windows security update changes a default that could knock some enterprise PCs offline. If admins never set the BlockNtlmSSO registry value, Windows now switches from Audit to Enforce mode and blocks NTLMv1-derived single sign-on credentials, which can break Wi-Fi, Ethernet, and VPN logins that still lean on older MS-CHAPv2-based setups in domain environments. Microsoft says this is by design, not a bug. Admins can set the value back to Audit (0) as a short-term workaround, but Microsoft warns that weakens protection, and NTLM is headed for removal in a future Windows release.
Drupal patches two critical remote code execution flaws
Drupal has fixed two critical bugs, CVE-2026-108097 in the core Field UI module and CVE-2026-108099 involving the contributed Entity List module, that could let attackers run code on sites using any supported version of the CMS. Neither can be exploited with Drupal core alone, but contributed modules or custom code open several paths, including at least one that needs no login. The security team had warned admins ahead of time that exploits could show up within hours or days. Sites should move to Drupal 10.6.19, 11.3.19, or 11.4.10 and Entity List 3.0.10; anyone who can’t update can uninstall the Field UI module as a stopgap. Drupal powers more than a million websites, many of them at universities and nonprofits.
Update: Deno has about a year left at Cloudflare, and Ryan Dahl is moving on to Workers
Last week we noted Cloudflare was buying Deno, the company behind the JavaScript runtime from Node.js creator Ryan Dahl. Here’s what changed: Dahl told The Register the deal is essentially an acqui-hire for him and 12 colleagues, who will work on expanding Cloudflare’s Workers platform for possible open-source use. Cloudflare will keep shipping bug fixes and security patches for the Deno runtime for one more year and then stop developing it, while Dahl looks for a new home for the MIT-licensed project. Dahl now says that choosing Node.js for building servers is “doing it wrong,” and that years of chasing Node compatibility left Deno as just another Node implementation.
Open-weight models now take nearly a quarter of spending on Vercel’s AI Gateway
Vercel says open-weight models handled most of the token traffic on its AI Gateway for the first time in August, and processed more than seven in 10 tokens by September. Their share of spending jumped from about one in 20 dollars in June to nearly one in four in September, while the average price per token was about 47 percent lower than in December 2025. Claude Opus 5.5 still drew the most September spending at 11 percent, followed by GPT-6 Astra and DeepSeek V4.1 Flash. DeepSeek alone handled 58 percent of tokens for just 6.6 percent of spend, and almost a third of teams switched their main AI provider in September. The gateway only sees its own customers’ traffic, so it’s a slice of the market, not the whole picture.
An AI lab will fix your plumbing for free if it can film the repair to train robots
Data-collection startup HUMXN is launching a program in Minneapolis, Chicago, and Miami that covers plumbing, electrical, HVAC, air-duct, and carpet-cleaning jobs at no cost to the homeowner. In exchange, the technicians wear sensors that record video of the work so HUMXN can use it to teach robots how skilled trades get done, and the pros still get paid. The company picked three cities with very different climates to capture a wider range of real-world repairs. Founder Thomas Unise says robots need to learn from how people actually handle physical tasks, not just from text and images scraped from the internet.
E Ink’s Kaleido 4 promises brighter, more colorful e-paper
E Ink has announced Kaleido 4, the next generation of the color e-paper used in devices like the Kobo Libra Colour and Kindle Scribe Colorsoft. The company says it doubles the previous version’s color gamut volume, raises white reflectance by 15 percent, and boosts contrast by 50 percent, which should help with the dim, washed-out look reviewers complained about on Kaleido 3 screens. It can also play video at up to 18fps. The first device announced with it is Hanvon’s M10C II, a productivity tablet that isn’t expected in the U.S., so it may be a while before it reaches mainstream e-readers.
Japan’s false alert for a magnitude 9.9 earthquake wasn’t a hack
Phones, TV broadcasts, and loudspeakers across Japan warned of a magnitude 9.9 earthquake on Friday afternoon, bigger than any quake ever recorded. The actual quake was a 6.9 near Russia’s Kamchatka Peninsula, hundreds of miles away, and the Japan Meteorological Agency pulled the alert eight minutes later, the first time an emergency earthquake warning has ever been withdrawn. An agency official told The New York Times it wasn’t a cyberattack, and the agency is investigating; one theory is that a magnitude formula meant for nearby quakes was applied to a distant one. At least one train line paused automatically, but no damage was reported.
That’s the Lanbrew tech roundup for today. Check back tomorrow at 10am CT for the next one.
Comments
Post a Comment