The day's tech, sifted: Nov 11, 2026

Wed, Nov 11

What matters today: yesterday Anthropic priced its IPO at up to $1.4 trillion; today its bankers are weighing a higher price, and Bloomberg's sources say orders already run to multiple times the shares available. Nvidia fell another 5.47% as the Huang affair raised the odds of chip export limits passing Congress this year, while rival chip startups raised billions. DeepMind released AlphaGenome 2, and in security Microsoft patched 1,093 flaws in one month as researchers found the Grok Build coding agent exposing thousands of users' private apps.

AI / LLMs

  • What Anthropic's roadshow video disclosed: Simon Willison transcribed the investor video, in which co-founder Tom Brown says Claude is trained and served on all four major chip families (AWS Trainium, Google TPUs, Nvidia and AMD GPUs) and that Mythos got Claude running on AMD in three days with almost no human help, a port that used to take dozens of engineers months. In the same video Dario Amodei says cancer, Alzheimer's, heart disease and mental illness will be "relics of the past" in under a decade. Why it matters: chip independence is being sold to investors as a moat, and these are claims from a sales video, not audited results.
  • Qwen4-Max in preview: Alibaba teased "Qwen4 is coming", and AlphaSignal reports preview access to Qwen4-Max across its Token Plan, QwenWork, Qoder and the Qwen app, aimed at coding and agent work with 1M-token support. Why it matters: a new Qwen generation would put a Chinese lab back in the week's frontier launch cycle; no independent numbers yet.
  • How OpenAI wants you to use Codex: at the AI Engineer Code Summit, OpenAI's developer team admitted Codex's cloud side has been awkward and pitched the fix: local and cloud runs merging into one executor you hand tasks between, voice as a control plane for managing agent threads, and dots (its personal agents) as the place tasks start. Why it matters: the labs now compete on how you supervise agents, not only on the model underneath.
  • Agent teams become a product feature: Cursor's Projects goes multiplayer, putting a team and its coding agents in one workspace so agents coordinate before edits collide, and Google's Antigravity /swarm command turns one prompt into agents that plan, build and verify in parallel. Why it matters: once everyone runs several agents, the merge conflict becomes the bottleneck.
  • Apple tests a local-model app: Apple's Claris unit is testing Harbor, a Mac app for running AI models locally. Why it matters: a first-party way to run models on a Mac would make local AI a default option, not a hobbyist setup.
  • Also shipping: MiniMax's H3 Image generates and edits images up to 4K, ACE Music v1 tops Artificial Analysis's vocal music leaderboard ahead of Suno v6, and ExecuTorch WebGPU runs PyTorch models on the GPU in a browser. Why it matters: image, music and on-device tooling keep moving while attention sits on the frontier labs.

Security & Privacy

  • Patch Tuesday, four digits again: Microsoft fixed 1,093 vulnerabilities, 63 of them critical, including an Outlook spoofing bug (CVE-2026-124194) already under attack. That makes 5,100+ patches this year against roughly 100 a month a year ago, a jump Microsoft credits to AI finding bugs. Chrome fixed its sixth zero-day of the year, Firefox 160 closed almost 100 CVEs, and Apple shipped out-of-cycle iMessage fixes. Why it matters: AI is finding flaws faster than teams can patch them; Ivanti's advice is to prioritize the small exploited slice (217 of about 95,000 CVEs this year).

  • Grok Build leaks its users' apps: security firm Claroc found that a design decision in SpaceXSI's vibe-coding agent lets anyone enumerate thousands of live shared projects despite randomized links, exposing business plans, medical exam results and employee IDs, plus 24 apps built by SpaceXSI itself. Why it matters: "unlisted" sharing is not access control, and apps built by prompt inherit whatever the platform got wrong.

  • Railway: the thieves wanted AI keys: the cloud host's report on its October hack says attackers stole customer secrets to resell AI provider keys, with unauthorized spend on the stolen keys estimated at about $740K (full report). Why it matters: model API keys are now loot worth breaching a hosting provider for.

  • A backdoor in fewer than 10 neurons: Microsoft Research's Lazarus attack reprograms dormant neurons so an agent behaves normally, and passes standard safety evals, until a hidden trigger (the example: the date rolling into 2027) makes it run attacker code. It needs only write access to a checkpoint, no retraining; reliable detection means diffing against the original weights. Why it matters: model files become a supply-chain risk like any binary, so provenance and tight agent permissions are the defense.

  • Ring cameras that move: Amazon announced the Ring Rover Cam and Ring Drone Cam, a patrolling robot and an indoor drone due from late 2026 and early 2027, both of which map the home's floor plan. Why it matters: a roaming camera with cloud processing collects far more of a home than a doorbell does.

  • $129M deepfake fraud arrests: New York federal prosecutors say three men, arrested in Spain and the UK, used AI to impersonate the voices and faces of company executives and law-firm attorneys, then talked employees into wiring money for supposedly sensitive deals. The alleged take between 2024 and 2026: more than $129M. Why it matters: a voice or a face on a call no longer verifies anyone, so payment approvals need a second channel.

  • TV makers questioned on tracking: The Verge sent every TV and streaming-device maker 17 yes-or-no questions about what their sets record, store and send to the cloud, starting from the premise that all of them collect viewing data. Why it matters: side-by-side answers give buyers something a privacy policy does not.

Devtools & Infra

  • AMD's Zen 6 server chips land: Tom's Hardware's first look at EPYC Venice, built on TSMC's N2P process, finds a dual 9996 up to 93% faster than the dual Zen 5 it replaces, and the 9686F often ahead of Nvidia's Vera CPU. Why it matters: agent workloads lean on CPUs as well as GPUs, and AMD has a generational lead to sell.

  • CoreWeave beats and raises: Q3 revenue rose 168% to $3.66B with a $132.5B backlog and higher 2026 guidance. Why it matters: days after its bond yields signaled credit stress, the demand side of the AI buildout still looks intact.

  • Cloudflare picks a leaderless consensus algorithm: Meerkat, its experimental service for keeping the global control plane's state in agreement, runs on QuePaxa, a 2023 algorithm, instead of Raft or Paxos. Those depend on one leader, and when the leader's link degrades every write stalls; Cloudflare cites a 2020 switch failure that left its dashboard 80 times slower for six hours. Why it matters: a rare production bet on a new consensus design, made because internet links are less predictable than a data center's.

  • IBM's Nighthawk r3: the new quantum processor is designed to run 10,000 operations across 520 qubits. Why it matters: operation count, not qubit count alone, decides how complex a problem a quantum chip can attempt.

  • GitHub's agent workflows go GA: Agentic Workflows are generally available with isolation and budget controls. Why it matters: sandboxing and spend caps are becoming the entry ticket for running agents in CI.

Startups & Industry

  • Anthropic's IPO, oversubscribed: a day after setting a $70 to $78 range, bankers are weighing a higher price, which the FT says could lift the valuation target to $1.5 trillion or more. By afternoon Bloomberg's sources said the book holds multiple times the shares available, with a Dario Amodei and Jamie Dimon livestream set for Friday. Anthropic also opened an Australian retail offering of Class A shares, and Data Center Dynamics reports another $4.3B in compute agreements with Microsoft. Why it matters: terms that came in under the hype yesterday look unlikely to stay there.
  • Nvidia, day three: shares fell 5.47% after Monday's 4% drop. Politico's sources say the affair has significantly raised the chances of Congress writing chip export limits into this year's defense bill, and Axios reports the Commerce Department received a national security warning about it. Semafor, reading 42 Huang interviews since 2022, finds his rhetoric grew more admiring of China and more critical of export rules after the first chip restrictions. Why it matters: the scandal is turning into policy, the outcome Nvidia has lobbied hardest against.
  • OpenAI's numbers leak: documents seen by the WSJ show Q3 revenue up about 50% on the quarter to $10B and adjusted operating margin improving from -150% to -109%. Why it matters: growth is fast and losses are narrowing, but OpenAI still spends about two dollars per dollar earned as its rival goes public.
  • Chip startups raise billions: London's Fractile raised $720M at a $7B valuation, per FT sources, up from about $1.1B in May, and Tenstorrent announced a $1.5B Series E at $6.2B. Why it matters: investors are funding alternatives to Nvidia in the same week its export fight worsened.
  • Deals: Oura is buying Counsel Health for $625M to pair ring data with an AI chatbot backed by in-house physicians, OpenRouter says it is joining Stripe, and Bloomberg's sources say Applied Intuition, which makes self-driving simulation software, is in talks to raise at least $1.5B at about $30B. Why it matters: buyers are paying for AI attached to a distribution channel (a wearable, a payments network), not for models alone, and at the top end rounds now come in billions.
  • Kalshi warned: the CFTC has told the prediction market it may face imminent charges for failing to notify the regulator about certain products. Why it matters: after a run of state court losses, its federal regulator is now a threat too.

Research

  • AlphaGenome 2: DeepMind's successor to last year's genome model predicts how a DNA mutation plays out in specific cell types, trained on 100M+ cell profiles across 68 tissues and open to non-commercial research. Why it matters: nearly every cell carries the same DNA but switches different genes on, so the cell-level view is where a variant's link to disease shows up.
  • Token spend is not productivity: JetBrains researchers linked five weeks of coding-agent usage to the commits of 404 engineers and found doubling token spend bought about 10% more commits, while the heaviest users spent 39 to 240 times the lightest for similar output. Why it matters: companies budgeting or ranking engineers by tokens are measuring usage, not results.
  • Simple agents hold up: a controlled test across 13 enterprise procedures found multi-agent and skill-augmented designs cost 1.4 to 4.7 times the tokens of a single agent for gains that were rarely significant. Separately, FlowPlan has agents plan repairs on a program's control-flow graph before touching code, cutting tokens by up to two thirds on SWE-bench Verified with Haiku 4.5 at the same fix rate. Why it matters: both point at architecture discipline, not more agents, as the cheap win.
  • Agents struggle to stop: DeputyBench tests whether enterprise agents check permissions before acting; across 12 models they were safe 88% of the time overall, but ran restricted actions about three times as often when permission evidence was missing as when it was explicitly denied. Why it matters: an agent that obeys a clear "no" may still act when it simply cannot verify a "yes".

Elsewhere

  • Pentagon hedges on launch: the Office of Strategic Capital lent Rocket Lab $150M two weeks after the same for Stoke Space, as SpaceX moves to wind down Falcon. Why it matters: the military is paying to keep alternatives alive before its workhorse rocket retires.

  • Washington: Senate Democrats opposed a lame-duck FCC nominee and plan probes of Chairman Carr once they take Congress in January, and Bloomberg reports at least 11 NSA and Cyber Command personnel died by suicide in the year to September, drawing White House and congressional questions. Why it matters: oversight of the agencies that run US communications and cyber policy is about to change hands.

  • AI ads in the midterms: the Wesleyan Media Project counted about 300 AI-generated or AI-enhanced political ads backed by roughly $187M in spending, 90% of the tracked ones for Republican candidates. Why it matters: synthetic imagery is now routine campaign material, though Poynter notes campaigns stayed wary of voter backlash.

Hacker News

The biggest tech story on the slice of the front page in hand is a post claiming a Google engineer pointed 48 AI agents at NASA lidar data and found an unmapped Maya pyramid in 72 hours; that is the author's account, not a confirmed find. Security and privacy filled much of the rest: VulnCheck's update tracking Anthropic-discovered vulnerabilities fits the Patch Tuesday entry above, and Mozilla wrote up research on making Firefox users harder to fingerprint.

On the tooling side, Microsoft announced .NET 11. Two hardware-flavored hacks drew crowds: one author used voltage glitching to unlock the A100 GPU inside a mining card, and a Show HN recompiles Black Ops II into a native ARM64 Mac app with no emulator, its author crediting Claude Opus 5.5 for the conversion tool. The most commented story had nothing to do with software: Toyota is reviving Scion, starting with a kei truck built in Indiana.

Threads

  • AI on both sides of the bug count: Patch Tuesday's 1,093 fixes and VulnCheck's tally of Anthropic-found flaws (Hacker News) show AI discovering vulnerabilities at volume; Lazarus and Grok Build show AI systems adding new ones.
  • AI assets as loot, AI output as bait: Railway's attackers wanted provider keys to resell and Lazarus needs only write access to a checkpoint, while the $129M fraud arrests and the midterm ad count show synthetic voices and faces already working at scale.
  • The agent bill: the JetBrains token study, the single-agent result and FlowPlan (Research) all question spend, just as Cursor Projects, Antigravity /swarm and OpenAI's Codex pitch push teams toward more agents; GitHub's GA answer is budget controls.
  • Listing week economics: Anthropic's book is oversubscribed and its Microsoft compute bill grew, OpenAI's leaked margins show the cost of that growth, and CoreWeave's backlog shows where the money lands.
  • Life after Nvidia: Anthropic's four-chip-family pitch, the Fractile and Tenstorrent rounds and AMD's Venice numbers all describe an industry reducing its dependence on one supplier, as the Nvidia entry moves from stock drop to export limits in the defense bill.
  • Verification by machine, claims by video: AlphaGenome 2 ships with a research license and data counts, while the roadshow's three-day AMD port and disease timeline remain company statements; the Maya pyramid post (Hacker News) sits in the same unverified column.