Silicon Valley Daily
Nscale announced $3.36 billion in convertible financing, with funding split between closing and an expected November tranche. Microsoft is bringing chat, coding and persistent agents into one Copilot app, while Meta is explaining the cloud workspace behind Muse. The D.C. Circuit upheld Anthropic’s defense procurement designation, and new research documented publicly readable databases and flawed agent evaluations. Cash schedules, usage bills and access permissions are all part of putting these systems to work.
Appeals court upholds Anthropic’s defense supply-chain designation
A divided D.C. Circuit panel rejected Anthropic’s challenge to the Pentagon’s supply-chain designation, preserving restrictions on Claude in defense work. The 2–1 ruling concerns a separate legal authority from the designation rejected by a San Francisco court; Anthropic says it is considering further review.
Microsoft brings Home, Code and Autopilot into Copilot
Microsoft’s redesigned Copilot combines chat and Office work in Home, app building in Code, and persistent cloud agents in Autopilot, formerly Scout. Home and Code are rolling out through Frontier, while Autopilot expands to private preview at month-end; longer-running agent work uses usage-based billing.
Nscale raises $3.36 billion in pre-IPO convertible financing
Nscale announced a $3.36 billion convertible-note financing led by Third Point, with $2.36 billion at closing and $1 billion from Nvidia expected in mid-November. The notes convert when its IPO completes, with Nvidia receiving non-voting shares; the proceeds support its AI cloud and data-center expansion.
Meta opens up Muse’s cloud workspace as downloads climb
Meta says Muse users can inspect files and run software inside their own cloud workspace, with sensitive controls kept outside that runtime. Sensor Tower estimates the app reached 3.4 million downloads by Thursday as Meta expanded promotion; other tracking firms report different totals.
UpGuard finds 16,326 Supabase databases with publicly readable tables
UpGuard identified 16,326 Supabase databases exposing readable tables; over half had schema indicators of personal information, rather than every row being individually inspected. Supabase told TechCrunch that projects have secure defaults and that customers control their configurations, placing access policies at the center of the dispute.
Transluce traces agent hacking attempts through public web-service logs
Transluce’s report traces AI-like traffic back to March and identifies attempted intrusions against three public-data services while agents sought ordinary facts. It links two targets to an OpenAI-attributed swarm but finds no successful exploit in the visible logs; OpenAI says its broader review may take months.
Irregular links several agent incidents to one faulty evaluation setup
Irregular’s CTO told The Verge that unintended internet access and a simulated company name matching a real domain caused several labs’ agents to target real systems. He says the incidents shared one evaluation scenario and that the company has tightened access controls, monitoring and pre-test checks.
Kiteworks recommends precautionary shutdowns after a threat warning
Kiteworks told customers to consider a precautionary shutdown after law enforcement warned of a possible imminent attack. Its security chief told TechCrunch that no compromise was known and that release 9.5.1 fixed known flaws, while the warning also covered potential unknown vulnerabilities.
Sony and Universal challenge Suno’s new model in another lawsuit
Sony and Universal allege that Suno’s v6 carries forward copyrighted material through outputs and distillation from earlier models. Suno has said the model uses a new dataset that includes user creations; the labels’ claim that this remains infringing is an allegation before the court.
Bitget revises wallet-incident losses to $387.5 million
Bitget’s latest notice raises affected assets to about $387.5 million after adding previously uncounted Zcash and TRON transfers. The exchange says the revision reflects a fuller tally of the original incident, and that Mandiant and SlowMist are assisting its investigation and recovery work.
Lightspeed targets a $250 million India fund focused on early-stage AI
An investor letter reviewed by TechCrunch sets a $250 million target for Lightspeed India Partners V, with commitments covering 80% of that amount. The planned fund is half the size of its predecessor and uses a shorter investment period focused on early-stage AI in India and Southeast Asia.
Pentagon budget request seeks $30.3 million for AI-assisted polygraphs
MIT Technology Review examined a Pentagon request allocating roughly $30.3 million across fiscal 2027–2031 to Polygraph Next. The program would study non-contact sensing and AI scoring for personnel screening; the budget request and planned validation work do not establish that the technology reliably detects lies.
Gravity Linux brings an experimental accelerated desktop to M4 Mac mini
Gravity Linux’s early alpha boots a Fedora-based desktop with GPU acceleration on the M4 Mac mini, using development work its authors say benefited from coding agents. The release targets developers: suspend, USB-C display output and Thunderbolt/USB4 are not ready, and the team warns of stability limitations.
Raspberry Pi revenue rises 90% as advance memory purchases help supply
Raspberry Pi reported first-half revenue of $256.9 million and 4.2 million boards shipped, up 90% and 17% respectively. The company credits earlier memory purchases with supporting supply and profit, while warning that the benefit from cheaper inventory has moderated.
Google tests Gemini business calls for eligible US Pixel users
Google is gradually opening Call for Me to eligible adult US Pixel 11 users with a paid AI plan and the Phone app beta. Gemini can ask about stock or manage bookings using the user’s number, while disclosing its AI role and letting the user monitor or take over the call.
Claude Code expands cloud sessions with credits for eligible subscribers
Claude Code cloud sessions can keep running on Anthropic’s infrastructure when the user’s computer is off. BleepingComputer reports promotional credits of $100 for eligible Pro subscribers and $250 for Max, limited to cloud sessions and accounts already subscribed when the September 23 promotion began.
Tesla details Semi production plans and charging requirements
Tesla used its Nevada Semi event to outline a factory target of up to 50,000 trucks annually and charging-network plans. Executives also described changes from pilot vehicles, including lower weight and roll-down windows; customer deployment still depends on charging infrastructure and fleet readiness.
Robert O’Callahan leaves Google over the pace of AI development
Google DeepMind engineer Robert O’Callahan wrote that he resigned because his chip-design work would accelerate AI in ways he no longer supported. He described his risk assessment as uncertain and plans to keep maintaining debugging tools, framing the decision as his own professional and ethical judgment.
Noctua and Forced Physics research quieter micro-channel air cooling
Noctua and Forced Physics are exploring whether JouleForce micro-channel cooling can fit desktop PCs and workstations. The research focuses on generating or reducing the high air pressure required without excessive noise; no product or release date has been announced.
DAWO builds an open blueprint for Dutch government workspaces
DAWO brings government, businesses and the open-source community together around a digitally autonomous workplace for the Netherlands. Its blueprint combines replaceable components for operating systems, cloud services and collaboration, including a NixOS-based workplace rather than a completed nationwide migration.
CISA adds WSO2 and Adobe flaws to its exploited-vulnerability catalog
CISA added CVE-2026-5430 in WSO2 products and CVE-2026-71362 in Adobe Commerce and Magento to its KEV catalog after observing exploitation. The agency directs federal civilian agencies to prioritize remediation based on exposure and risk, while encouraging other organizations to do the same.
Elementor fixes a flaw requiring an administrator to open a malicious link
Patchstack disclosed a CSRF flaw in Elementor 4.3.0 and 4.3.1 that could make a logged-in administrator perform an attacker-chosen action after opening a link. Version 4.3.2 fixes the issue; the report does not mean every Elementor installation is vulnerable to the same attack.
DriveWealth breach exposes historic records of some Revolut customers
DriveWealth notified customers of a September 4–5 intrusion involving historic brokerage records, including some former Revolut customers. The broker said passwords and payment details were not affected, while Revolut said its own systems were not breached; the number of affected customers remains undisclosed.
US seeks to intervene in X’s challenge to a European fine
The US Justice Department applied to intervene in X and Elon Musk’s challenges to the European Commission’s €120 million Digital Services Act fine. Washington argues that the liability theory reaches too far across corporate boundaries; the filing supports an appeal and does not itself overturn the fine.
Tower plans a $4 billion optical-chip manufacturing hub in Japan
Tower Semiconductor’s CEO told Nikkei that a planned $4 billion investment would make Japan its largest optical-communications chip production hub. The expansion targets the optical links used in AI servers, adding manufacturing capacity through upgrades and planned new facilities.
Meadow ships a small companion phone with maps, rides and music
Meadow has begun shipping its small Android companion phone, retaining services such as maps, ride-hailing, music and tickets while leaving out social feeds and browsing. TechCrunch’s hands-on report found it useful outside the home, but users need a separate phone number and subscription rather than simply mirroring their main number.
GitHub shows how canvases turn agent chats into shared tools
GitHub’s latest canvas guides show how a Copilot session can build a board, checklist or dashboard that both the user and agent update. The interface can be saved as an extension, giving repeated workflow actions a visible set of controls instead of requiring a new chat instruction each time.
GitHub details the performance gains from moving to CSS Modules
GitHub’s engineering write-up explains its staged migration from runtime CSS-in-JS to CSS Modules, using feature flags and visual comparisons to preserve behavior. It reports that an earlier Primer migration cut server rendering time by 55% and component initialization time by 25%, with later work removing remaining styling dependencies.
a16z introduces Cosign as a directory built around professional endorsements
In its newly posted discussion, a16z introduced Cosign as a directory of people and startups organized around public professional endorsements. The speakers described categories for mentors, people one would work with again and emerging talent, presenting the system as a way to make trusted networks more discoverable.
UK backs heat networks using waste energy from industry and data centers
The UK has announced £90 million for new and upgraded heat networks, with waste heat from data centers and industry among the potential sources. One London scheme initially plans electric-barge transport before a later tunnel connection; the government describes projected benefits and planning-dependent infrastructure, not completed service.
CommonGrid organizes public US utility and grid data for reuse
CommonGrid combines public utility, grid-operator and geographic-boundary information in an open dataset and explorer. Its repository lists 3,132 utilities and provides JSON and GeoJSON files under the ODbL license, with source documentation intended to make reuse and corrections easier.
Microsoft mitigates a black-screen issue on some Windows virtual desktops
Microsoft says some Windows 11 systems can fail to load the desktop after August’s preview update and later releases, particularly Azure Virtual Desktop hosts using FSLogix. A Known Issue Rollback is available while a permanent fix is being prepared; the issue differs from an earlier wallpaper-only bug.
Researchers validate AI-assisted solutions to two archived Enigma messages
TechCrunch reports that cryptanalysts have validated solutions to two previously unsolved archival Enigma messages using OpenAI’s Astra and Anthropic’s Opus. The accompanying Crypto Cellar accounts show different amounts of human guidance, making these specific historical codebreaking results rather than a general Turing-test result.