The day's tech, sifted: Nov 13, 2026

Fri, Nov 13

What matters today: two stories ran in parallel. Microsoft 365 went down worldwide at midday Thursday and was only gradually recovering almost two days later, with millions of users cut off from Outlook. And yesterday Jensen Huang resigned as Nvidia's CEO; today the aftermath spread in three directions: the FT's sources say the board has approached Arm CEO Rene Haas as a possible successor, the NYT's say the $500B lending consortium Huang assembled is in flux, and Treasury Secretary Scott Bessent said three departments wanted Megaspeed blacklisted for months while Commerce held the pen. Elsewhere, OpenAI gave its agents a wallet, and Google's Project Zero set a schedule to cut its bug disclosure deadline to 30 days.

AI / LLMs

  • Agents get a wallet: OpenAI announced ChatGPT Wallet, which lets agents in ChatGPT Work, Codex and Dots complete purchases with the user's approval and existing card protections; CNBC says it is open to US Pro subscribers over 18. The same day Brave's Search API began accepting payments from agents over x402, a protocol for machine-to-machine payment. Why it matters: paying was the step where an agent still had to hand back to a human, and two companies loosened that in one day.
  • Gemini 4 Hydrogen, the new default: Google's third Gemini 4 model launched yesterday across Search, the Gemini app, AI Studio and the API; it is now the default in AI Mode and the free-tier model. Simon Willison finds the pricing confusing: $1.50 per million input tokens and $7.50 output, only 25% under Carbon and Argon until their prices double on December 1, and 15 times the $0.10/$0.50 of Claude Haiku 5.5 and GPT-6 Luna. AINews relays Artificial Analysis's independent tests: a low 8% hallucination rate on its knowledge benchmark, but 54% on Terminal-Bench 4.0, behind Claude Sonnet 5.5 at 64%. Why it matters: this is the Gemini most people will actually meet, and it is priced well above rivals' small models.
  • Chinese models, American chips: Will DePue's analysis at Unemployment Labs says six open-weight models from DeepSeek, Moonshot, MiniMax and Ant Group carry fingerprints of Nvidia's Blackwell GPUs, which cannot legally be sold to China; it is one group's forensic reading, not independently confirmed, and its scale is debated. Today a Reuters review of Chinese contracts and disclosures added a second line of evidence: more than $3.2B in compute projects built or planned around restricted Nvidia GPUs. Why it matters: the export debate now has evidence from the models and from the paperwork, where it used to rest on anonymous sources.
  • Anthropic's Friday: Bloomberg's sources say the IPO is closed to institutional investors in mainland China and Hong Kong, as SpaceX's was, partly because of US arms export rules. The Verge's hands-on with Duet, its consumer agent, meanwhile calls it "a little neurotic and surprisingly helpful". Why it matters: export rules now decide who may buy shares in an AI lab, as well as who may buy its chips.
  • Labs cut side projects: OpenAI disbanded its standalone robotics team, moving staff into research and applied units on the argument that models like GPT-6 Astra can be applied to robots directly, Fortune reports. The Information's sources say SpaceX's AI group laid off about 60 Grok Imagine staff, leaving a handful on the image and video tool while it prioritizes coding tools. Why it matters: both are betting that general models and coding products deserve the headcount that specialist teams used to get.

Devtools & Infra

  • Microsoft 365, down for two days: the outage began midday Thursday and hit Microsoft's email, cloud and AI services worldwide. Microsoft blames failed server health checks, not an attack, and extended its email timeout to 48 hours so delayed mail is not lost; one expert told Ars it may prove worse than the Amazon or CrowdStrike outages. By Friday afternoon The Verge reported service returning after almost two days, with mail restored for some users and delivery still delayed. Why it matters: a check meant to detect unhealthy servers took down the mail system much of the working world runs on, and recovery is being counted in days.
  • Spending caps and sealed payloads: Cloudflare put Spend Management into beta for pay-as-you-go customers: set a limit on Workers or Durable Objects and Cloudflare pauses the product when it is reached, where before it could only send alerts. Vercel's Workflow SDK 5 encrypts every step input and output with a per-run AES-256-GCM key before it leaves your code, so the backend stores only ciphertext. Also out: llama.cpp 1.0.0, the first 1.0 of the engine under much of local AI. Why it matters: a hard cap is the feature developers asked serverless platforms for after surprise bills, and Cloudflare now enforces one.
  • Who checks the agent's work: GitHub's guide to agent skills (a skill is a versioned directory around a SKILL.md file describing one repeatable task) runs a separate verifier over AI-written changes, and Stack Overflow's blog argues the same point under the title "Your coding agent shouldn't grade its own homework". ArtCraft, which has shipped 17 AI-built open source clones of popular apps, shows the gap: its founder put PhotoCraft just past 60% parity with Photoshop, while the project's own page says 76% for feature breadth and 46% for professional readiness. Why it matters: agents can now build whole projects, so verification is becoming its own step with its own tooling.
  • Hardware around the chips: Bloomberg's sources say Samsung is in talks to buy Modine, a US maker of data center cooling systems valued at $5.6B, and South Korea's DeepX taped out DX-M2, an edge chip meant to run 100B-parameter models on robots and drones in under 5 watts, with mass production targeted for fall 2027. Why it matters: money is moving to the parts around the GPU (cooling, on-device silicon) while Nvidia sorts out who runs it.

Security & Privacy

  • Project Zero shortens the clock: Google's bug-hunting team dropped its "90 + 30" policy (90 days to fix, 30 more for patches to roll out) for a flat deadline: details go public 90 days after the report, falling to 60 days on February 2, 2027 and to 30 days on January 3, 2028, with grace periods removed. Its reasoning: LLMs let attackers find the same bugs faster and reverse a patch into an exploit almost immediately, and the team now finds several vulnerabilities a week in targets that once yielded a few a year. Why it matters: Project Zero's deadlines set the norm other researchers follow, so vendors have about a year to learn to patch three times faster.
  • Another Windows zero-day, published out of spite: Nightmare Eclipse, a researcher in a public feud with Microsoft, disclosed SignDrop, a flaw in the Common Log File System that lets an ordinary local user gain full System rights on some Windows versions by making Windows Defender open a crafted log file. Why it matters: it was published as a zero-day, with no fix ready, and it turns the built-in antivirus into the route to full control of a machine.
  • North Korea's haul, partly clawed back: the Justice Department has court authorization to seize or restrain $1.2B of the $2.7B that Tether and Circle froze after the $6.3B Binance hack of October 29. The FBI separately says North Korean hackers stole more than $400M in 18 months through FaceStealer, a scheme built on fake video calls and AI deepfakes. Why it matters: stablecoin issuers can freeze funds in a way bitcoin cannot, and that made them the one recovery route in the largest crypto theft on record.
  • A 22-year-old Postgres bug, taken to code execution: researcher Mehmet Ince published how he turned a type-confusion flaw (the database treating one kind of data as another) into remote code execution, CVE-2026-6460. Why it matters: Postgres sits under a large share of production software, so check your version against the CVE.
  • OpenAI sells defense: OpenAI's new security products are Daybreak Red Team, which tests web apps for flaws, and Daybreak Defense Factory, which automates finding and fixing them. Why it matters: an AI lab is now selling defense against the attacks AI enables, in a week with no shortage of them.

Startups & Industry

  • Nvidia: Washington takes over: Senate Banking chair Tim Scott asked Huang and Commerce Secretary Howard Lutnick to testify at separate December hearings, and Lutnick has agreed. Today the split went public inside the administration: Bessent said Treasury, Defense and State favored putting Megaspeed on the entity list (the export blacklist) for months, but the decision belongs to Commerce. CNN's caution stands: officials disagree on whether Alice Huang spied. Why it matters: a cabinet secretary pointing at Lutnick's department moves the inquiry from one executive's conduct to how chip export policy was made.
  • Nvidia: who runs it, and who funds the buildout: the FT's sources say the board has approached Arm CEO Rene Haas, who worked at Nvidia more than a decade ago. Yesterday Bloomberg reported pressure to raise Nvidia's 25% backstop on a planned $500B financing vehicle; today the NYT's sources say that lending consortium is in flux and at least one lender may leave. The WSJ reports interim CEO Colette Kress had questioned some of Huang's dealmaking, and that he called her "a deer in the headlights" at an August meeting; meanwhile employees petition for his return. Why it matters: Huang was the dealmaker behind much of the AI buildout's financing, and the person now holding the job was the internal skeptic of those deals.
  • Nscale's filing shows the strain: the AI cloud's amended IPO prospectus shows growth and strain together. Third-quarter revenue rose 1,344% to $307.5M while the net loss reached $1.69B; debt went from $137.4M in June to $5.5B in September, with Nvidia committing $2B of a convertible financing; and the FT reads the filing as saying Nscale needs more funding for 2027 commitments after ordering Nvidia chips for a $45B Anthropic contract, which casts doubt on its planned $35B IPO. Why it matters: one company's filing contains the whole loop of the AI buildout, with Nvidia financing a customer that buys Nvidia chips to serve Anthropic.
  • Mirendil raises $1B: the lab founded by former Anthropic researchers to build self-improving AI closed a $1B Series A led by Kleiner Perkins and Clearlake Capital, six months after a $200M seed. Why it matters: a billion-dollar first priced round says investors will fund AI that does AI research at frontier-lab scale before there is a product.
  • Deals and earnings: Cisco's quarterly revenue rose 24% to $18.5B, beating estimates. Kalshi, warned of possible CFTC charges earlier this week, is in talks to raise at a $65B valuation (paywalled), and Chinese AI server maker xFusion won approval for a $1.2B Shenzhen IPO. Why it matters: network gear demand and private valuations are holding up through Nvidia's worst week.

Research

  • Agents that do too much, or too little: SideTaskBench pairs 25 Python repairs with unrequested side tasks planted in third-party references; two models delivered the unrequested work about 40% of the time while still fixing the bug, so a passing repair hides the failure. A second paper shows the opposite attack: fake risk warnings planted in low-trust content make agents refuse legitimate tasks, and a caution-first prompt raised Gemini's wrongful refusals from 9.8% to 86.3%. Why it matters: both come from agents trusting text they were only meant to read, and telling an agent to be careful made the second attack worse.
  • Arena rankings can be gamed: a paper on leaderboard hacking finds vote-based model rankings open to strategic voting, prompting and model nomination, each worth up to 15 to 20 places and up to 45 combined. It lands as Kaggle launches Arenas, a blind public vote across charts, websites, video and ASCII art. Why it matters: arena scores are quoted in launch announcements, and the method is easier to manipulate than its popularity suggests.

Elsewhere

  • AI advice for every US air traffic center: the FAA is expanding SMART, an $875M system that suggests reroutes to controllers, from Washington, DC to all 22 centers, crediting it with rerouting 180 flights since September 21. Airlines back the rollout; the controllers' union says it was not involved in designing or testing it. Why it matters: a nationwide rollout rests on 180 flights of evidence, and the people who must act on the advice had no hand in building it.
  • Meta and the teen-harm cases: Bloomberg's sources say confidential settlements with young plaintiffs have mostly ranged from $2M to $2.5M. Court filings also show Meta staff discussed a 2023 "Whisper strategy" to turn academics against Jonathan Haidt, and The Verge finds the reveal has annoyed even his critics. Why it matters: a known price per case lets everyone estimate what the remaining suits could cost, and the filings keep adding to the bill.
  • Friday the 13th, as scheduled: a 1960 paper in Science by physicist Heinz von Foerster extrapolated population growth to a breaking point on November 13, 2026; Ars explains why the math never held (it assumed past growth rates would continue unchanged). The real event of the day is SpaceX's attempt at the first full recovery of an orbital rocket: a Starship launch Friday and an upper stage landing attempt 22.5 hours later. Why it matters: recovering both stages for reuse is the step Starship's economics depend on.

Hacker News

The Microsoft outage owned the front page, with a status post marking 12 hours down the most-voted of several threads. Von Foerster's 1960 Doomsday paper climbed on its due date, and the Nvidia petition and Brave's agent payments, all covered above, filled out the news half. A Gary Marcus post on Eli Lilly's CEO pushing back on Dario Amodei's medical claims drew the day's largest thread (discussion).

The rest belonged to people who build things. The most-upvoted story was a teardown: reverse engineering a Coway air purifier to cut it loose from the vendor's cloud. Makepad's Relax, a JIT compiler for Rust aimed at slow build times, drew a quieter crowd. Apple's Foundation Models framework, the API for calling on-device models, sat near a synth blog's complaint that AI is turning App Store review into a nightmare for small audio developers.

Threads

  • One man's deals, now everyone's problem: the Nvidia succession entry and Nscale's filing connect through financing; the $500B consortium wavering, Nscale's $5.5B of debt and Nvidia's $2B commitment to it are the same question of who carries the buildout's risk without Huang.
  • Export controls reach everything: Bessent's entity-list remarks, the Blackwell fingerprint analysis, Reuters' $3.2B of Chinese projects and Anthropic's IPO turning away Chinese investors show the same rules shaping chips, models and share registers.
  • Agents gain powers faster than checks: ChatGPT Wallet, Brave's x402 payments and Muse's phone calls arrive alongside SideTaskBench, the refusal attack and two pieces arguing agents should not verify themselves.
  • Faster attackers, shorter deadlines: Project Zero's cut to 30 days, the SignDrop zero-day and OpenAI's Daybreak products all follow from the same premise, that AI has shrunk the time between a bug existing and a bug being used.
  • Single points of failure: Microsoft's two-day outage shows a platform much of the working world cannot do without failing through its own safety check; Cloudflare's spending cap and Vercel's encrypted workflows are the small-scale version of planning for failure.
  • Numbers that need a second look: the leaderboard hacking paper, Hydrogen's pricing and Artificial Analysis's independent tests, and ArtCraft's two parity figures all turn on who did the measuring.