Silicon Valley Daily
Anthropic released Claude Haiku 5.5 on Wednesday at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, a tenth of Haiku 4.5's list price, and halved Sonnet 5.5 cache-read pricing to $0.10. OpenAI shipped Intelligent UI, which puts interactive diagrams, calculators and tappable elements inside ChatGPT answers, to paid tiers with GPT-6, while Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens an Unacceptable Risk after more than 4,000 test prompts and OpenAI published its first teen usage report the same morning. Biohub, the Department of Energy, NIH, Google DeepMind, Isomorphic Labs and Meta described a $1.8 billion mix of funding, data and compute for the Virtual Biology Initiative; Google opened its SynthID watermark detector to everyone and launched Playground, a prompt-driven game maker for U.S. adults; Microsoft priced the Surface Laptop Ultra at $2,599 and demonstrated a Copilot that acts on local files. Elon Musk said Grok Bot will route tasks to Claude Opus 5.5, Midjourney and Suno. CNBC reported Apollo and banks are in talks to finance SpaceX's $40 billion Nvidia purchase, Nous Research raised $90 million at a $1.2 billion valuation per the Wall Street Journal, and the FBI and Secret Service warned that FortiBleed credential attacks are locking organizations out of their own Fortinet firewalls.
Anthropic ships Claude Haiku 5.5 at a tenth of Haiku 4.5's price
Anthropic released Claude Haiku 5.5 on Wednesday, pricing input at $0.10 and output at $0.50 per million tokens for prompts up to 100,000 tokens, versus $1.00 and $5.00 for Haiku 4.5; prompts above 100K cost $0.50 and $2.50. The company says the model costs around 75% less to run on average, is its fastest model to date, is the first Haiku with an adjustable effort setting, and scores 72.4% on OSWorld 2.1 and 39.2% on Terminal-Bench 4.0, against 15.7% and 0.0% for Haiku 4.5. It also halved Sonnet 5.5 cache-read pricing to $0.10 per million tokens, which it says makes Sonnet about 20% cheaper on most agentic work, and introduced a monthly API credit for Max and Team subscribers. Haiku 5.5 is available on the Claude Platform, AWS, Google Cloud and Azure; its cyber safeguards permit more defensive tasks than Sonnet 5.5's but still block penetration testing.
OpenAI rolls out Intelligent UI, putting interactive visuals inside ChatGPT answers
OpenAI began rolling out Intelligent UI on Wednesday with GPT-6 to Pro, Plus, Business and Enterprise users, with free and Go tiers following Thursday. Instead of text alone, ChatGPT generates tappable buttons, task-specific calculators, interactive charts and editable diagrams when it judges them useful; product manager Aarush Selvan demonstrated a prompt about how an airplane wing generates lift that returned a set of diagrams. In a pre-launch test, Wired's Reece Rogers generated a slug anatomy diagram with clickable parts, a San Francisco apartment affordability calculator with sliders, and a seat-map explorer comparing four Alaska and Delta aircraft. Users can tell ChatGPT to produce fewer visuals. Wired compared the feature to the generative UI Google added to Search earlier this year.
Common Sense rates ChatGPT for Teens an Unacceptable Risk as OpenAI posts first teen usage report
Common Sense Media's Youth AI Safety Institute said Wednesday it tested more than 4,000 prompts on accounts registered to 13- to 17-year-olds, before and after the August 18 launch of ChatGPT for Teens, and rated the product an Unacceptable Risk, urging OpenAI to restrict ChatGPT to adults until the gaps are fixed and independently tested. Testers spent up to an hour discussing suicidal ideation, self-harm or disordered eating on newly created parent-linked accounts and the linked parents received zero alerts; the Institute also said the chatbot missed more than one in four warranted crisis referrals, that study mode surfaced a Show me the answer option, and that adult-registered accounts never switched to the under-18 experience even after testers said they were 13. The Institute is funded in part by the OpenAI Foundation and says it keeps editorial independence. An OpenAI spokesperson told Axios the company does not believe the testing accurately reflects how the safeguards work in practice. The same morning OpenAI published its first ChatGPT for Teens progress report: teens spend under 15 minutes a day on average, fewer than 2% use it for more than three consecutive hours, nearly 1.2 million teens used Learning Visualizations in one week, and a College Planner is coming for U.S. students in grades 10 to 12.
Biohub, DOE, NIH, DeepMind, Isomorphic and Meta outline $1.8B Virtual Biology Initiative
Biohub said Wednesday that it, the U.S. Department of Energy, the National Institutes of Health and new funding partners are committing $1.8 billion in funding, data, computation and measurement technology to build open, AI-ready biological data for predictive models of cells. The figure combines more than $500 million from DOE over five years through its Genesis Mission, datasets NIH will coordinate from more than $500 million in prior federal investment, $300 million from Google DeepMind, Isomorphic Labs and Meta, and Biohub's own $500 million announced in April. The Allen Institute, Broad Institute, Gladstone Institutes, Human Cell Atlas, Human Protein Atlas and Wellcome Sanger Institute are joining the effort. According to Reuters as relayed by The Decoder, Biohub science head Alex Rives said commercial funders get one year of exclusive access to the data they paid for before it is released, government-funded work carries no such restriction, and a first dataset should be ready in about a year.
Google opens SynthID Detector to everyone; it checks only supported watermarks
Google on Wednesday opened its SynthID Detector website to the public, letting anyone upload an image, video or audio file to check whether it carries a SynthID watermark from Google or partners OpenAI, Nvidia and Kakao, with Apple support coming. Google says more than 180 billion images and videos and 240,000 years of audio have been watermarked since 2023, and that verification built into Search, the Gemini app and Chrome handles over a million requests a day. Ars Technica reports the site requires a Google, OpenAI or Apple login and allows roughly 10 checks per user per day, a limit Google engineers said deters attempts to build watermark-removal tools. The detector reports only whether SynthID was used to create or edit the media; content from models without SynthID, including open-weight models, will not be flagged, and the site carries a disclaimer to that effect.
Google launches Playground, a prompt-to-game platform for U.S. adults, with Unity Spark to follow
Google Labs launched Playground on Wednesday, an experimental browser-based platform where users describe a game in text and the system builds it, with no coding required. Creators pick a genre or start from scratch, choose 2D or 3D and single- or multiplayer, upload visuals to be turned into assets, and can keep games private, share a link or publish to an Explore gallery with leaderboards; multiplayer is supported in select genres. Playground is available to U.S. users aged 18 and older, browsing and playing are free, and generation runs on a weekly token system with higher limits for Google One AI subscribers. A Google spokesperson told The Verge it runs on Gemini, Nano Banana and Lyria. Google and Unity also announced Unity Spark, a professional-oriented tool with Asset Store access; Google says it is in testing with a closed beta coming soon. The launch page reached the Hacker News front page.
CNBC: Apollo and banks in talks to finance SpaceX's $40B Nvidia GPU purchase
Apollo Capital Management and several banks are in discussions with SpaceX to help finance its purchase of $40 billion of Nvidia GPUs, a person familiar with the talks told CNBC, a day after the Financial Times first reported the financing. The deal would rely primarily on the investment-grade debt market and would likely use the GPUs as collateral, with Apollo in a leading role; the source said talks are preliminary and lenders are not finalized. SpaceX raised $25 billion in bonds two weeks after its mid-June IPO, but CNBC notes AI-related bonds have since sold off and credit spreads have widened, and that fixed-income investors are demanding higher yields. A credit-industry source said the market still assumes GPUs hold their value for about seven years. Apollo declined to comment.
Musk: Grok Bot will route tasks to Claude Opus 5.5, Midjourney and Suno
Elon Musk said on X on Tuesday night that SpaceX will use the best back-end model for any given task in Grok Bot, naming Claude Opus 5.5, Midjourney, Suno and other leading APIs, with simple questions routed to small, fast models and complex ones to large models. The Information's Grace Kay reported the move as SpaceX's AI unit using competitors' models to power Grok Bot, a shift from relying on its own Grok models alone. The post drew commentary that SpaceX is positioning Grok Bot as an orchestration layer across third-party models rather than a single-model assistant.
Microsoft demos Hybrid Intelligence Copilot that acts on local Windows files
At its Windows and Surface event on Wednesday, Microsoft showed a Copilot upgrade it calls Hybrid Intelligence, in which apps mix local and cloud models and the assistant gains access to files on the PC and the ability to take actions across the operating system. Copilot EVP Jacob Andreou played a video demo in which the Autopilot agent, asked to send an accountant what she needs for taxes, searched folders for documents, renamed them, zipped them and drafted an email with the archive attached. Andreou said Hybrid Intelligence features will come to Copilot over the next couple of months. Microsoft also showed a new Windows search bar that performs quick actions such as toggling dark mode, adjusting mic volume with an inline slider or sending a text, due this fall on Windows 11 according to Windows and Surface chief Pavan Davuluri.
Surface Laptop Ultra ships Oct. 16 from $2,599; RTX Spark Dev Box priced at $5,999
Microsoft said the Surface Laptop Ultra, built on Nvidia's Arm-based RTX Spark chip with up to 128GB of unified memory, launches October 16 and is available for preorder now. The base model costs $2,599 with an 18-core CPU, 24GB of RAM and 512GB of storage; the top configuration is $5,899.99 with a 20-core CPU, 128GB of RAM and 1TB. The 15-inch HDR touchscreen reaches 2,000 nits, and the laptop introduces Magnetic Connect, a magnetic charging cable that snaps into a standard USB-C port. Microsoft also opened preorders for the Surface RTX Spark Dev Box at $5,999, a 128GB unified-memory mini PC with a 100-watt thermal envelope aimed at running local models exceeding 120B parameters, shipping in November with Visual Studio Code, Git, GitHub Copilot, WSL, Python and Node preinstalled under its Project Zenith developer setup. The Verge notes the Dev Box costs more than Nvidia's own DGX Spark amid RAM shortages.
WSJ: Nous Research raises $90M at $1.2B valuation on 22M Hermes downloads
Nous Research, maker of the open-source Hermes AI agent, has raised $90 million at a $1.2 billion valuation to expand into enterprises, the Wall Street Journal's Yuliya Chernova reported. The company says Hermes has been downloaded more than 22 million times since February. Posts on X linked from Techmeme cited a $1.5 billion figure and roughly $36 million in annualized revenue by mid-September; the Journal's headline valuation is $1.2 billion and the revenue figures were not in the reported article summary.
Personal agents: Tab exits stealth at $300M valuation, Wired finds Dots buggy, a16z counts the payers
Tab, a personal assistant operated over iMessage and WhatsApp, said Tuesday it had emerged from stealth at a $300 million valuation; founders Brennan Erbz, Stafford Schlitt and Ammar Amdani declined to disclose the funding amount, and investors include SV Angel, Valar Ventures and American Spirit. Amdani says the agent has its own phone number, computer and wallet, that user data is never used for training, and that nothing irreversible happens without a yes. Wired's Reece Rogers spent several days with OpenAI's $100-a-month Dots agent: it got his name wrong, offered to solve a captcha it could not, misheard a mumble and replied that it loved him too, yet produced a usable shortlist of couches. In a16z video discussion of its seventh Top 100 Consumer AI Apps report, partners Olivia Moore and Josh Elman describe personal agents as the biggest new trend; the report itself, published October 5, adds a first spending ranking and finds 4.5% of U.S. consumers held a paid ChatGPT, Gemini or Claude subscription in August. Tony Fadell told TechCrunch the first wave of AI gadgets failed because they solved no real need, and that a trusted agent will have to run on-device.
Decision models arrive: OpenAI's Decisions API beta and Liquid AI's open-weight d1
OpenAI opened a public beta of its Decisions API, a dedicated endpoint that evaluates text or images and returns a yes/no probability, a pick from predefined categories or a scale score, which the company says runs about ten times faster than the Responses API. Only gpt-6-luna is supported, priced at $0.10 per million input tokens with no output or cache charges, though regional processing premiums and long-context multipliers apply; zero data retention and HIPAA use are available to eligible customers, and general availability is expected in the coming weeks. The Decoder also reports OpenAI cut its paid API tiers from five to three. The same day Liquid AI released two open-weight decision models on Hugging Face: d1-3B, which it says scores 48.57 on Decision Index 0.2.1 ahead of Decider 35B-A3B at 47.11 and answers a question in 16 milliseconds on an Nvidia Jetson AGX Thor, and an experimental d1-omni-600M that handles text with images or audio. Both answer in a single forward pass rather than generating tokens; benchmark figures are the vendors' own.
Crunchbase: North American startup funding hit $92B in Q3, $61B of it to AI
North American startups raised $92 billion in the third quarter, down 35% from the second quarter but up 50% year over year, according to Crunchbase News' Joanna Glasner. AI companies took $61 billion of the total. The quarter saw 11 acquisitions of $1 billion or more, while IPO activity remained quiet, and funding came in well below the all-time peak.
FBI and Secret Service: FortiBleed credential attacks are locking admins out of Fortinet firewalls
The FBI and U.S. Secret Service issued a joint advisory dated October 6 describing FortiBleed as an active global credential-compromise campaign against internet-facing FortiGate firewalls and SSL VPN gateways. Attackers use credentials from earlier breaches and infostealer logs for credential stuffing and password spraying, extract password hashes from compromised devices and crack them offline on GPU clusters, then create new accounts and in some cases delete or change passwords on existing ones, locking organizations out while attempting lateral movement. The advisory cites SOCRadar's count of 86,644 compromised devices across 194 countries and links the campaign to initial access brokers supplying ransomware affiliates; The Register reports evidence pointing to INC/Lynx and Payload affiliates. The agencies advise restricting internet-facing management access, terminating active admin and VPN sessions, resetting credentials and enabling phishing-resistant MFA. The advisory describes credential abuse, not a newly disclosed Fortinet zero-day.
Parallel Systems raises $100M Series C for autonomous battery-electric freight rail
Parallel Systems, founded in 2020 by former SpaceX avionics engineers, raised a $100 million Series C led by AVP with Hillspire, Agility Global, Cobalt Capital, Anthos Capital, Congruent Ventures, Riot Ventures and Collaborative Fund participating, bringing total funding past $200 million. Its Panther rail vehicle carries several tons of freight up to 500 miles without an operator and runs in coupler-free platoons that split without human intervention, targeting the roughly 60% of U.S. freight trips under 500 miles that trucks dominate. CEO Matt Soule says the company won Federal Railroad Administration approval about a year and a half ago to operate near the Port of Savannah, where it has run on 160 miles of track to prove safety, and that its first commercial payload is coming up very soon. The round funds manufacturing of the third-generation vehicle.
Meta adds LLM detection for ads that covertly lead to child abuse material; 33.2M pieces actioned in H1
Meta said Wednesday it has deployed new measures against child sexual exploitation, including a large language model system to detect signposting, meaning ads that look benign but are suspected of directing people to illegal material or harmful activity off-platform, plus analysis of where an ad sends users, retrospective sweeps, a red-teaming AI agent that probes its own defenses, and improved detection of banned users returning on new accounts. Meta said it actioned 33.2 million pieces of child sexual exploitation content on Facebook and Instagram from January to June 2026, more than 97% found before user reports, including 5.3 million in India at over 98% proactive. The counts cover pieces of content, not unique offenders, and are Meta's own figures. TechCrunch notes the announcement follows Meta's August agreement to pay up to $18 billion to settle a child-safety suit brought by 29 U.S. states.
Healthleap raises $38M for AI that flags hospital patients for clinical review
Healthleap, founded in South Africa in 2022 by siblings Jemima and Josiah Meyer, raised $38 million: an $8 million seed co-led by Sequoia Capital and First Round Capital and a $30 million Series A led by Hummingbird Ventures, TechCrunch reported. The platform reads electronic health records nightly, extracts mentions such as weight loss or trouble swallowing from clinicians' notes with language models, combines them with labs and vitals, and writes a risk score into the care team's workflow each morning; it does not diagnose, only highlights patients for review. CEO Josiah Meyer says it is deployed in more than 50 hospitals including Penn Medicine, Cedars-Sinai, Intermountain, Houston Methodist and Emory, screening for malnutrition and delirium, with programs for aspiration pneumonia, pressure ulcers and heart-failure readmission still in clinical validation. Revenue growth and return-on-investment figures are the company's own.
Lumen: PoeLLM malware hit 3,400+ exposed AI servers, finds its C2 through a poem on GitHub
Lumen's Black Lotus Labs said a cryptomining campaign it calls PoeLLM, active since at least April 2026, has affected more than 3,400 servers with peak activity exceeding 800 active servers a day, predominantly in the United States and Western Europe. Victims mostly run exposed open-source AI services such as LiteLLM and Ollama, plus the Gotenberg PDF converter and Gitea, with possible Ivanti Sentry targeting. The malware derives its command-and-control address from keywords in a poem hosted in a GitHub repository, mapping words to numbers; the operator has changed the poem 11 times. It deploys XMRig and Iron miners connected to the Kryptex mining service and turns victims into scanners that probe ports 3000 and 4000 and attempt to exploit LiteLLM flaw CVE-2026-42271, which BleepingComputer says can be chained with CVE-2026-48710 for unauthenticated remote code execution. Lumen assesses with moderate confidence that the operator is Italian-speaking and has blocked traffic to the C2 servers; BleepingComputer corrected its initial 2,100-server figure to 3,400 after the researchers updated their report.
Xreal prices Android XR Aura glasses at $1,279, $20 under Meta's VR Glasses, shipping this year
Xreal opened preorders for its Aura glasses on Wednesday at $1,279 for 12GB of RAM and 256GB of storage and $1,499 for 16GB/512GB, initially for customers who paid a reservation deposit, with general preorders opening in the coming weeks and shipments to the U.S., Canada, Japan and South Korea; 9to5Google reports a first delivery wave before the holidays and more markets in 2027. The glasses weigh under 95 grams and tether to a puck running Android XR with Google Play and Gemini on Qualcomm's Snapdragon Reality Elite, with a 70-degree field of view, hand tracking and Bose audio. Wired's Boone Ashworth, who tried both, found Meta's $1,299 VR Glasses smoother for seated use and typing, while Aura's optical see-through lenses make it easier to walk around; Meta's glasses do not launch until early next year.
Finland orders Google to halt two data center sites until environmental assessments finish
Finland's Licensing and Supervision Agency on Tuesday ordered Google subsidiary Tuike Finland Oy to suspend construction work at planned data center sites in Muhos and Kajaani until environmental impact assessments are complete, CNBC reported. The suspension covers tree removal, topsoil stripping, excavation, quarrying, ditching and road building, must take effect no later than October 23, and does not cover planning, measurements or soil surveys. A Google spokesperson told CNBC the company had fallen short of its own high standards, had acted in good faith under the Forestry Act, and will follow the agency's guidance, adding that it plans tree planting across 130 hectares at Muhos. The order comes a month after Google committed $15 billion to AI infrastructure in Finland through 2028, its largest investment in Europe.
Chips: Musk says Terafab will run its own fab as Intel reaffirms 14A deal; Qualcomm to license Huawei's LogicFolding
Intel CEO Lip-Bu Tan told Bloomberg on Wednesday that Intel remains part of Terafab, the chipmaking venture of SpaceX, Tesla and xAI that in April agreed to use Intel's 14A process. Elon Musk rejected rumors that TSMC would take over the project, writing on X that his companies will build and run the fab and that TSMC might at most sublease part of it; Tom's Hardware argues a shared fab would raise process-flow, IP-isolation and liability complications. Separately, Tom's Hardware relays a report that Qualcomm will license the patents behind Huawei's LogicFolding, a two-die hybrid-bonded design already used in the Kirin 9050 Pro; a Geekerwan teardown shows the lower die holds mostly cache and I/O while compute sits on top, and Huawei's own papers claim up to 66% lower NPU power at the same performance and a density increase from 155 to 238 million transistors per square millimeter. Huawei does not name a process node.
Florida, Iowa, Nebraska and Montana sue TP-Link over router security and China ties
Florida Attorney General James Uthmeier announced on October 6 that the state is suing TP-Link, alleging the company misrepresented how secure its routers are, how separated it is from China and what could happen to customers' data; Iowa, Nebraska and Montana filed similar suits. Uthmeier claimed Chinese state-sponsored hackers exploited TP-Link routers in American homes, that many models cannot update automatically and are no longer supported, and that less than 1% of the parts at TP-Link's Vietnam plant come from Vietnam. TP-Link has said its U.S. business separated from its Chinese parent in 2024 and has built routers in Vietnam since 2018, and The Register reports the company rejects the allegations. The suits follow federal inquiries opened in 2024 and an FCC import ban on new consumer routers for which TP-Link has sought a conditional exemption.
Krebs: suspected ShinyHunters leader detained while extorting Boeing spin-off Jeppesen ForeFlight
KrebsOnSecurity reports that the teenager known as Rey, identified as Saif Al-din Khader of Amman, Jordan, was detained while ShinyHunters was extorting Jeppesen ForeFlight, the navigation unit Boeing sold to Thoma Bravo for $10.55 billion in November 2025; Reuters reported on October 3 that he is cooperating with the FBI. Boeing acknowledged the extortion claims, and Jeppesen ForeFlight said its investigation found no impact to operations or products. Krebs says Rey took over the ShinyHunters brand after Dutch police arrested Pepijn van der Stap on September 15, and that the group mass-exploited Oracle PeopleSoft flaw CVE-2026-35273, first as a zero-day in June, to steal data from dozens of organizations including an FBI recruitment site that exposed records on more than 5,000 personnel. Reuters reported October 5 that the FBI removed an Accenture contractor over the unpatched site.
Nvidia: fine-tuned Nemotron reached gold level at both IOI 2026 and IMO 2026
Nvidia said specialized versions of its Nemotron 3 models reached gold-medal level at both the International Olympiad in Informatics and the International Mathematical Olympiad this year. Nemotron-3-Ultra-CC, a 550-billion-parameter model with 55 billion active parameters, scored 535.4 of 600 at IOI 2026 using supervised fine-tuning and an iterative generate-evaluate-refine loop called GenCorrect, above the 361.12 gold threshold and the top human score of 498.27; Nvidia says the run was live under contestant constraints but unofficial and not in the official ranking. For IMO 2026, a generate-verify-refine system built from Nemotron 3 Ultra checkpoints scored 30 of 42, above the gold threshold of 29, with proofs graded by official IMO graders. Nvidia trained the coding specialists on 22,000 curated problems and the math SFT model on 414,890 filtered examples, and is releasing models, data and recipes on Hugging Face.
curl pulls 8.23.0 forward to Oct. 14 to fix 22 vulnerabilities, one rated HIGH
curl maintainer Daniel Stenberg said the project will ship curl 8.23.0 on October 14, a few weeks earlier than planned, after receiving one vulnerability report that highlighted what he called a rather significant flaw. The release will fix 22 security vulnerabilities, including CVE-2026-92392, which the project rates HIGH; curl has published only two HIGH-severity CVEs since 2021, the most recent being CVE-2023-38545. No details will be made public before the European morning of October 14, when the fixed version and updated Rock-solid curl builds ship; the distros list and paying support customers will be notified ahead of time. Affected versions, trigger conditions and exploitation status have not been disclosed.
CTA: fully U.S.-made consumer tech would cost $185B to $230B and 555,000+ added workers
The Consumer Technology Association estimated in research shared with members this week that reaching full U.S. production of ten common household product categories by 2031 would require $185 billion to $230 billion in capital expenditure, 555,000 to 668,000 additional full-time employees, more than double the current U.S. computer and electronics manufacturing workforce, and 19.1 to 19.5 billion kilowatt-hours of electricity a year, Ars Technica reported. Smartphones would cost 152% more to make, laptops 93% more, smartwatches 97% and TVs 41%, with weighted average retail prices rising 27% to 55% if companies pass on a quarter to half of the increase. The CTA proposed reshoring only final assembly as a first step, at $16 billion to $19 billion, and urged eliminating tariffs on components from trusted trading partners.
Stuut raises $52.5M Series B led by Insight for AI order-to-cash agents
Stuut, which uses AI agents to automate enterprise order-to-cash cycles including invoice collections, raised a $52.5 million Series B led by Insight Partners, bringing total funding to $93 million, SiliconANGLE and Axios reported. The company says it serves more than 150 clients.
Gardens Interactive raises $35M+ Series B at $190M pre-money ahead of first launch
Gardens Interactive, a studio led by makers of Sky: Children of the Light and Journey, raised a Series B of more than $35 million led by Lightspeed at a $190 million pre-money valuation, taking total funding to $71 million, Bloomberg's Cecilia D'Anastasio reported. The studio is building a social role-playing world and is still running pre-alpha playtests ahead of its flagship launch.
Utah pilot lets Nolla Health's AI prescribe acne treatments with phased physician review
A one-year pilot in Utah will let Nolla Health's $4.99-a-month app examine patients with mild-to-moderate acne and recommend medication from a list of eight approved topical treatments, TechSpot reported citing Bloomberg. Users scan their face and complete a medical history; the AI scores acne severity and skin oiliness and selects a prescription. Oversight is phased: the first 100 prescriptions are reviewed by physicians before reaching a pharmacy, the next 400 are sent by the AI and reviewed retrospectively each week, and after that a doctor reviews at least 10% of prescriptions monthly. CEO Luis Wenus said side effects of the approved medications are limited to localized irritation or dryness. The Hacker News headline described the program as prescribing without human oversight; the reported structure retains physician review at each stage.
Researchers tie an agent fleet on Tencent Cloud to 1,810 daily scans of Alibaba's Amap
A preliminary report from a research group calling itself the Swarmchasers says a fleet of AI agents running on Tencent Cloud behind a proxy named hysandbox-ats used the urlquery.net sandbox service to pull entrance-share data from Alibaba's Amap mapping service for parks, zoos, museums and hospitals, reaching 1,810 scans on October 4 alone and 2,048 scans across 216 locations through that date, Tom's Hardware reported. The agents generated Alibaba anti-bot tokens, borrowed public API keys and loaded Alibaba's own Baxia anti-bot scripts; 211 runs were labeled claude, but the researchers say the code matches Chinese models rather than Anthropic's. Fifteen of 16 readable result inboxes on webhook.site were created from Tencent Cloud, 13 by script. The HY prefix suggests Tencent's Hunyuan models, though no public documentation of hysandbox exists and nothing ties the fleet directly to Tencent's own hysandbox domain.
GitHub: credential-bearing pushes rose 2.59x in nine quarters; ML push protection in private preview
GitHub said screened pushes rose 2.84x and pushes containing credentials rose 2.59x over nine quarters, with no statistically detectable trend in the prevalence of secrets per push, arguing that developers are being outpaced by volume rather than becoming more careless. The company described a ModernBERT-based classifier that evaluates candidate batches in under two milliseconds to catch secret formats that regex patterns miss. Push protection using the classifier is in private preview, with broader access planned later this month; some post-push scanning surfaces update earlier, and AI credits may apply. The figures are GitHub's own telemetry.
Ring's first smart lock recharges by hand-crank; $249, U.S. sale planned for early 2027
Amazon's Ring announced its first smart lock, which includes a power dial that recharges the lock by hand if the battery dies, plus four ways in: fingerprint reader, backlit keypad, the Ring app, or Alexa. It connects over Amazon Sidewalk rather than home Wi-Fi, fits a standard single-cylinder deadbolt, and founder Jamie Siminoff told TechCrunch the battery should last about a year. Bloomberg reports a $249 price. Ring says the lock will be available to purchase in the U.S. in early 2027 in black, silver and brass, and its disclosure notes the device has not yet received FCC authorization and cannot be sold until it does. Ring also introduced new 4K cameras alongside the lock.
Nobel Prize in Chemistry goes to Kagan and Soai for asymmetric autocatalysis
The Royal Swedish Academy of Sciences awarded the 2026 Nobel Prize in Chemistry to Henri B. Kagan of Université Paris-Sud and Kenso Soai of Tokyo University of Science for the discovery of non-linear effects and autocatalysis in asymmetric organic synthesis. Kagan showed in 1986 how to produce a greater excess of one mirror-image molecule than previously thought possible; Soai designed in 1995 the first reaction with the potential to be homochiral and in 2003 demonstrated a reaction that formed only one of two possible mirror images. The 12 million Swedish kronor prize is shared equally. The committee said the work explains how homochirality can emerge spontaneously and has guided reaction design in pharmaceutical manufacturing.
Wired: Pentagon's Tradewinds fast-tracks AI buys from OpenAI, Anthropic and Google via 5-minute videos
Wired reports that Tradewinds, a procurement program run by the Defense Department's Chief Digital and AI Office since late 2022, lets companies pitch AI products in videos of five minutes or less; accepted pitches gain post-competitive status that satisfies competition requirements and, a DOD official said, has enabled awards in under a week in several cases. OpenAI, Anthropic and Google are listed as participants and declined to comment. This year's focus areas include tools to enhance force lethality and agents to assist the joint targeting process including kill chain execution; a prior Biden-era topic on responsible AI guardrails is gone. The status eases use of other transaction agreements, which can reach $500 million without notifying Congress; Senator Joni Ernst told Wired reviews have flagged questionable costs and weak oversight, and her Stop Secret Spending Act will require public reporting within three years.
Googlebook's Better Together phone features only work with Pixel at launch; Samsung support in weeks
Google has confirmed that Better Together features on its new Googlebook laptops, including app streaming, phone file access, notification mirroring and Continue On, do not currently work with Samsung phones even those running Android 17, Ars Technica, 9to5Google and Android Police reported. Early buyers said devices such as the Z Fold 8 appear to pair but do not stay connected. Google says an update enabling Samsung support will roll out in the coming weeks, but only for Samsung phones already on Android 17; Motorola and OnePlus are still finishing Android 17 updates, leaving Pixel owners as the only users with full integration today. Ars notes Google had marketed the laptops as engineered to work with your Android phone out of the box.
Wired: nonprofit run by SpaceXAI's general counsel pays staff detailed to the National Design Studio
A Wired investigation found that at least two members of the National Design Studio, the Executive Office of the President unit led by Airbnb cofounder and former DOGE member Joe Gebbia, are employed by the America by Design Foundation, a recently formed nonprofit that appears to be run by James Burnham, a former DOGE member and current general counsel of SpaceXAI. Documents reviewed by Wired show Amittai Groot and John Cafazza were detailed to the Office of Personnel Management under the Intergovernmental Personnel Act on part-time terms, reporting to Christina Bonarrigo, who is now NDS chief of staff. It is not clear who funds the foundation or how much money it has; the individuals and Bonarrigo did not respond to requests for comment. University of Michigan professor Don Moynihan called not knowing who pays these people troubling.
Photopea developer says GitHub has not removed ad-stripped clones a month after his report
The developer of Photopea, a browser-based photo editor, wrote on Hacker News that people are asking AI models to take the site's JavaScript, strip the ads and republish the result on GitHub as a new product, that tens of such repositories exist, and that users have emailed him complaints about bugs that turned out to come from the clones. He says he reported the repositories to GitHub on September 4, 2026 and they have not been taken down a month later. The post reached the Hacker News front page.
Dutch tax office abandons Microsoft 365 cloud rollout for on-premises and European open source
The Netherlands Tax and Customs Administration has dropped its planned Microsoft 365 cloud migration in favor of services on infrastructure it controls plus European open source tools, according to a letter to parliament from State Secretary for Finance Eelco Eerenberg reported by The Register. Email and calendars move on-premises in 2027, with storage and collaboration following later that year and in 2028. The rollout, selected in 2025, had reached about 5,000 of 47,500 employees when the Dutch Advisory Council on ICT Assessment recommended in June that it be rolled back, saying the public-cloud design fell short on information security, information processing and future-proofing, citing confidential documents transiting American servers and reliance on standard Microsoft encryption rather than Double Key Encryption.
COSMIC bans AI-generated contributions as a GNOME developer argues for accepting AI bug reports
System76 has updated the contributor guidelines for its COSMIC desktop to require a declaration that a pull request contains no LLM-generated content, including code, comments and descriptions, The Register reported. GNOME Calendar and GNOME Extensions already restrict AI-generated contributions. GNOME developer Michael Catanzaro, in a post titled The Era of Software Quality, or the Era of Ostriches?, argues the project must accept AI-generated bug reports because bot-found flaws are too valuable to ignore; he previously cut GNOME Security's disclosure deadline from 90 to 30 days effective August 1.
Platformer: at The Curve, speakers float a hard cap on how intelligent AI systems may become
Casey Newton reports that at The Curve, an annual conference in Berkeley for AI lab executives, nonprofit leaders, officials and journalists, multiple speakers under the Chatham House Rule said the industry may want to limit how intelligent a future system is allowed to become, which depending on scope could amount to a de facto ban on superhuman systems. Newton ties the urgency to fallout from the OpenAI-Hugging Face incident and to recent OpenAI and Anthropic posts on progress toward recursive self-improvement, and notes Anthropic CEO Dario Amodei has called for some kind of speed limit on that process. He adds that no lab or country can enforce such a cap alone and the current U.S. government opposes limits, while lab leaders and officials differ on how close danger is.