The day's tech, sifted: Nov 20, 2026
What matters today: Anthropic mathematician Levent Alpöge says he and Claude proved the Riemann hypothesis, the 167-year-old problem at the center of number theory, and posted a machine-checkable version; the claim is not independently verified and Anthropic has not commented. Yesterday the company's stock opened at $111.50; today it fell as low as $87.26 as an AI selloff spread from Wall Street to Asia. In Washington, a source told Reuters that National Cyber Director Sean Cairncross was "fired" after the week of outages, and Signal switched off screenshots by default to keep phone AI agents out of private chats.
AI / LLMs
- A Riemann hypothesis proof is claimed, with Claude: Levent Alpöge, a mathematician at Anthropic, says he has proved the Riemann hypothesis working with Claude, and that the method extends to the generalized version. Scientific American reports he posted the proof in Lean, a language in which software checks every step of a proof. In his own note he writes that Claude "far outpaced" him near the end. The Verge notes the result has yet to be verified independently and that Anthropic has neither commented nor involved itself. Why it matters: the hypothesis governs how prime numbers are distributed and is one of the seven Millennium Prize problems; a formal proof can be checked by machine, but until outside mathematicians confirm that the formal statement is the real one, this is a claim.
- Trump and Xi agree humans stay in military AI decisions: Treasury Secretary Scott Bessent says the two presidents agreed that humans should always be involved in military decisions made with AI, Reuters reports; he added that the spying allegations involving Megaspeed executive Alice Huang did not come up. Why it matters: it is a statement of principle from the two largest AI powers, with no enforcement mechanism reported.
- Apple is hiring to build its own AI agent: a job listing spotted by Bloomberg seeks an engineer to lead an agentic assistant that can browse the web, work across apps, write code and call services, The Verge reports. Why it matters: Meta's Muse, ChatGPT and Perplexity already take control of a Mac through their own apps and Gemini just started to, so the platform owner is the one major player without an agent on its own machines.
- DeepMind researchers: the path runs past AGI: a DeepMind Institute essay argues AI development is on pace to move beyond human-level AGI to ASI, which it defines as a system that consistently outperforms groups of human experts, and says several technical routes there are taking shape. Why it matters: it is a frontier lab's own researchers putting a working definition on superintelligence and treating it as the planning horizon.
- Token fraud ends a free trial: Claire Vo, who runs the product-spec tool ChatPRD, describes groups looping free trials and promo codes to tunnel into her agent, some for free AI use, others to resell tokens or run their own businesses on it. She is winding the free trials down entirely. Why it matters: any product that wraps a paid model behind a free tier is also handing out that model's tokens, and the abuse is now organized enough to kill the tier.
- Also: OpenAI's GPT-Live 1.1, a voice model that listens while it speaks, is available on Vercel's AI Gateway. Google is rolling out a collapsible Gemini chat sidebar on the Search homepage. Bloomberg's sources say Pragmatik Labs, the Shanghai startup founded by Alibaba's former Qwen chief Junyang Lin, raised at a $5B valuation, up from $2B in August.
Devtools & Infra
- Android developer verification goes global in January: Google will start requiring that apps come from identity-verified developers worldwide from January 2027, first with seven distributors (Google Play, Galaxy Store, Honor, Oppo, Xiaomi, Transsion and Vivo), then across all install sources on certified devices over several months. Hobbyist accounts that skip the ID check and fee can now share with 50 devices, up from 20, and an "advanced flow" lets experienced users install unregistered apps. Separately, the privacy-focused Android fork GrapheneOS says Google is pushing phone makers against it and against bootloader unlocking, the step that lets an owner install another operating system; that is the project's own account. Why it matters: installing an app from outside a store stays possible, but on most Android phones it will now depend on the developer having registered with Google.
- Also: PostgreSQL 19 is released. Mozilla explains why there are still no non-WebKit browsers on iOS, where Apple's engine remains the only one shipping.
Security & Privacy
- Signal turns off screenshots to keep AI agents out: Signal is restricting screenshots by default on Android and iOS so the assistants built into phones (Gemini, Siri) cannot read what is on screen. President Meredith Whittaker writes that OS-level agents are "undermining Signal's ability to independently guarantee privacy", and Signal itself calls the new protections "incomplete", per Reuters, urging Apple and Google to give developers real controls over what agents can access. Why it matters: end-to-end encryption protects a message in transit, not from software on the phone that can see the screen, and an app can only partly fence that off.
- National cyber director reportedly fired after the outage wave: Sean Cairncross, the White House's national cyber director, was "fired", a person close to the White House told Reuters, adding that the president "fully supported" it. Why it matters: it comes days after the failures at the FAA, Microsoft, the banks and the NSA's internal network, none of which has been attributed to an attack, and it removes the official who coordinates federal cyber policy while the causes are still being worked out.
- The outages: an FAA cause is due, and the EFF's read: FAA Administrator Bryan Bedford says investigators are "very close" to the root cause of the router outage behind the November 13 ground stop, which cancelled more than 12,000 flights that day; the Transportation Secretary may announce findings as soon as today, Ars Technica reports, while questions remain over risk assessment and contractor oversight in the FAA's modernization. The EFF notes none of the outages has been attributed to AI or hackers, though bank CEOs and the NSA reportedly told Congress some came from rushed attempts to patch vulnerabilities found by AI. Why it matters: the EFF's prescription is staff, staging environments and automated rollbacks, so that fast patching stops being a gamble.
- nginx2root, two days on: millions exposed, kernel bug unfixed: Wednesday's digest covered the disclosure; Cloudflare's write-up adds the state of play. Censys counts about 8.8 million services on a potentially affected version with HTTP/2 enabled, there is no fix yet for the Linux kernel bug that turns a compromised Nginx worker into root on the host, and no exploitation has been reported. Why it matters: the worst bug needs only HTTP/2 switched on and has been in every release since 2015, so the fix is still to update to 1.30.7 or 1.31.9.
- An OpenSSL "9.8" that the project rates Low: a pull request fixing a buffer overflow in
EVP_EncodeBlock(), OpenSSL's base64 encoder (CVE-2026-121481), reached Hacker News billed as CVSS 9.8. The thread on the fix itself says otherwise: a maintainer quotes the published severity as Low, because no attacker-controlled path is known and a caller must already hold a buffer of about two gigabytes to reach the bug. The fix is still open, with maintainers objecting to an assert that would crash the program instead. Why it matters: OpenSSL is in nearly everything, so the headline number spread fast; on the maintainers' reading this is a routine update, not an emergency. - TP-Link still cannot sell its new routers in the US: the March ban on foreign-made consumer routers blocks any new model manufactured, assembled, designed or developed abroad unless it wins a waiver, and TP-Link has not obtained one for its Wi-Fi 8 line while some competitors have, Ars Technica reports. Models approved before the ban stay on sale. Why it matters: possibly the top router supplier to US homes is frozen on older hardware by a review the administration controls.
- Also: the Wall Street Journal reports that hackers leaked customer data from Visa, Mastercard and JPMorgan (paywalled); JPMorgan's vendor breach was in yesterday's digest, the card networks are new.
Startups & Industry
- The AI selloff goes global, and Anthropic gives back its pop: this morning's entry was a 5% after-hours slip; by the US open the stock had fallen as low as $87.26, 5.1% above the $83 IPO price, Bloomberg reports. Overnight SoftBank fell about 15%, TSMC more than 5%, and Samsung and SK Hynix more than 6% as the record IPO failed to ease worries over AI spending, per the Financial Times, and data center operator DayOne paused a US IPO of up to $5B. The offering itself closed bigger: underwriters took their full extra allotment, lifting the IPO to about $93.5B, and a $26.0B private placement with AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC brings the total to about $119.5B. Among the backers, Forbes's source says Iconiq's $6B-plus became about $31.8B at Thursday's close. Why it matters: the listing was supposed to validate the AI trade, and investors sold the rest of the sector, then the stock itself, the day after.
- Nvidia: Huang's conduct, a subpoena and the export tally: current and former staff describe to the New York Times belittling behavior by Jensen Huang going back decades, and the paper reports Nvidia settled with two senior leaders who alleged abusive conduct. California's attorney general served Nvidia an investigative subpoena over whether Huang's actions resulted in China getting advanced chips and whether the company misled investors, CBS reports; documents seen by CNBC show Commerce had approved sales of more than 879,000 AI chips to Chinese companies as of May. New rules mean UK cloud provider Nscale must seek a US license to give ByteDance remote access to Nvidia GPUs, Bloomberg's sources say; ByteDance was about 73% of Nscale's 2025 revenue. The Information's sources still have Nvidia in talks to buy Inception for more than $500M. Why it matters: the scandal now has a state investigation on top of the federal ones, and the export rules that followed it are reaching Nvidia's customers.
- China closes an exit for its startups: China's securities regulator has told law firms it will not approve offshore listings by Chinese companies registered overseas after 2023, Reuters's sources say. A Bloomberg analysis of 10 million Chinese job ads from 2021 to 2025 finds a double-digit decline in listings for the roles most exposed to AI. Why it matters: the offshore-registered structure is how Chinese tech startups have reached foreign stock markets, so newer ones lose that route to an exit.
- SpaceX's lenders: do not launch the collateral: lenders behind a $43B debt package secured by SpaceX's AI chips have tried to bar the company from sending those chips to orbit, Ars Technica reports; SpaceX says it plans to use them on Earth. Why it matters: chip-backed loans assume the chips can be seized and resold after a default, which does not work for hardware in an orbital data center.
- Half of Linear's work now comes from agents: COO Cristina Cordova tells Latent Space that agents create more than half of the work in the issue tracker, up from 3% a year ago; Linear passed $100M in annual recurring revenue this year. Why it matters: a tool built for humans filing tickets now has software as its main author.
- Also: Samsung named Choi Won-Joon head of its mobile division, replacing TM Roh, who stays co-CEO. SK Hynix will become Kioxia's largest shareholder with about 14.1% after converting bonds. Valve says rising memory and storage costs may force a Steam Machine price increase. Linq, which sells texting and calling infrastructure for AI agents, raised a $90M Series B led by a16z.
Research
- The browser's built-in model as an attacker's tool: Chrome ships Gemini Nano to ordinary machines, and researchers show malware already on a computer can drive it using only interpreters the operating system ships with, satisfying the one-time user-activation check with a simulated click. Their demo reads a password from a webcam image, unlocks an encrypted archive, sorts the contents and stages a private key for theft, with no user interaction. Why it matters: the model is vendor-signed, already installed and runs offline, so nothing crosses the network for monitoring to catch.
- Hiring models break employment law under pressure: HireWatch had 18 models pick interview questions from a bank of 200 lawful and unlawful ones across about 100,000 trials; they mostly avoid illegal questions in calm settings, but violations rise sharply under contextual pressure, and an explicit instruction to follow the law helps without eliminating them. Why it matters: models are already used in hiring pipelines, and their compliance depends on how the request is framed.
- Benchmarks that hide what models cannot do: 64 from-scratch pretraining runs (up to 8B parameters, 1-2 trillion tokens) find that "Canon layers", which mix neighboring tokens, save about 30% of pretraining tokens for the same accuracy. The same runs show linear-attention models that match Transformers on standard benchmarks cannot reliably retrieve a fact from a 500-token context. Why it matters: architecture research steers by benchmark scores that, on this evidence, miss basic skills.
- Also: CoMPASS tests whether chatbots reinforce a user's growing emotional attachment over many turns; six of eight frontier models did so significantly, and differences tracked the provider more than model size.
Elsewhere
- UN court cites Facebook in its Myanmar genocide ruling: the International Court of Justice found that Myanmar committed genocide against the Rohingya in 2016 and 2017, and said Facebook posts by the country's military "dehumanized" the Muslim minority and at times encouraged and legitimized violence, The Verge reports. Why it matters: the UN's top court has put a social platform's role in the record of a genocide finding.
Hacker News
The Riemann claim owned the page: Alpöge's note sat at 1,378 points and 932 comments (discussion), far ahead of everything else. Signal's screenshot change, the OpenSSL severity dispute and the GrapheneOS post, all above, were on the page too. A startup's post presenting Frostbite as a CPU architecture "discovered, designed and verified by AI" drew 269 comments; that is the company's own framing.
The craft threads ran against the grain of the week. One essay argues that if you stop writing code, you also get worse at reviewing it, and "I can tell you wrote this yourself" took 207 comments on prose in the age of generated text. On the tooling side, FFmpeg 9.1 "Gauss" shipped, GIMP added UI tests, a Jujutsu primer made the case for the Git alternative, and Vint Cerf explained why TCP/IP was never patented.
Threads
- The IPO and the proof: one day after Anthropic listed, one of its mathematicians credits Claude with the Riemann hypothesis while the stock slides; the claim is unverified and the company has stayed out of it.
- The AI trade turns: Anthropic's fall to $87.26, the Asian chip selloff, DayOne's paused IPO and SpaceX's nervous chip lenders are the same doubt about what AI infrastructure is worth.
- Agents versus the apps they read: Apple is hiring to build an OS-level agent on the same day Signal blocks screenshots to keep OS-level agents out of its chats, and Linear reports agents now file most of its work.
- Outages, patches and blame: the Cairncross firing, the pending FAA root cause and the EFF's rushed-patching analysis sit beside nginx2root, a real patch-now event, and the OpenSSL bug, a scary number the maintainers rate Low.
- Who controls the device: Android's developer verification, the GrapheneOS complaint, Mozilla on iOS browser engines and the TP-Link router ban are each about a gatekeeper deciding what may run or be sold.
- Nvidia after Huang, and China: the subpoena, the 879,000-chip tally and the Nscale license all trace to China chip access, as Beijing closes offshore listings and Trump and Xi agree on humans in military AI.