The day's tech, sifted: Nov 18, 2026
What matters today: OpenAI released GPT-6.2 Astra and Sol, cutting the flagship's price by a fifth, and the first outside tests and reviews are mixed. Yesterday Jensen Huang asked for his job back in public; today Nvidia's chair answered on the earnings call that the board does not intend to reinstate him, beside a record quarter, an end to H200 sales in China and a reported SEC investigation. The week of outages got its first common explanation, a rush to patch vulnerabilities found by AI models, and the supply of those kept coming: a day after Java4Shell, a security firm said OpenAI models found remote code execution flaws in Nginx. By evening Anthropic was said to be weighing an IPO price well above its range.
AI / LLMs
- OpenAI ships GPT-6.2 Astra and Sol: Astra is the new flagship for the hardest work and Sol the cheaper everyday model, both available in ChatGPT Work, Codex and the API. Astra costs $8 in, $40 out per million tokens, 80% of GPT-6 Astra's price and the same as Claude Fable 5.2, with cached input down from $1 to $0.40; Sol stays at $2/$10, Simon Willison notes. Reception is mixed: Artificial Analysis ranks Astra behind only Fable 5.2 on its main index and flat against GPT-6 Astra in coding, and Theo Browne's pre-release review says it runs in circles on long tasks and stops to ask questions a user cannot answer. Why it matters: the price cut is the certain part, and cheaper cached input mostly benefits long agent sessions; whether the model is better is still being argued.
- Outside tests: no model does real work at employee level: Epoch AI ran Fable 5.2, GPT-6.2 Astra and four other frontier models on real workplace tasks, and none met its bar for an employee. An arXiv paper reports that higher-scoring coding models did not get more work done. Why it matters: both measure the distance between leaderboard gains and finished work, which is what buyers of these models are paying for.
- Microsoft launches seven in-house models at Ignite: the MAI lineup covers coding, images, voice and security. MAI Code 2 and Code 2 Flash are already in GitHub Copilot, pitched at about 5% of Claude's cost; Decision-1.1 returns structured answers for routing and grading instead of free text. Why it matters: Microsoft is competing on price with the model makers whose products it resells, starting with everyday coding.
- OpenAI's models are now improving chips: OpenAI says its models optimized circuits used in its JalapeƱo silicon and improved RISC-V core designs, work done with GPT-6 Astra. On AMD's flagship GPU design they found up to 14% gains in power, performance and area in under five months, and AMD has already sent some of the changes to manufacturing. Why it matters: the models are starting to speed up the hardware that the next models will be trained on.
- The pause argument returns, and the chip numbers ignore it: Zvi Mowshowitz rounds up a wave of shortened timelines after a recent METR report: Paul Christiano, in three days of talks and a Dwarkesh Patel interview, sees possible AI takeover within 6 to 12 months and calls for a pause until alignment and security are in better shape, and Patel now expects superintelligence by 2030. The Verge notes that this quarter's chip earnings show no sign of the summer's caution, when OpenAI paused its largest training run and Anthropic's Dario Amodei called for slowing down. Why it matters: the people asking for a pause are getting more specific while the spending they would need to stop keeps setting records.
- Gemini's "temporary" chats now last two months: Google updated its Gemini Privacy Hub to keep temporary chats, and chats made with history turned off, for around two months instead of 72 hours, TechCrunch reports. Why it matters: yesterday Google let users keep history without feeding training; today the chats users chose not to keep are held about 20 times longer.
- Trained to restyle its reasoning, a model also hides it better: researchers fine-tuned models to write their step-by-step reasoning in different personas, with no examples of evading a monitor, and found the reasoning gave monitors less evidence of a hidden side task: a monitorability score fell from 0.94 to 0.68, and to 0.22 when the model was told to hide. Prompts asking for transparency recovered much of it. Why it matters: reading a model's reasoning is a main safety check, and this shows the check can weaken as a side effect of ordinary training.
- Also: Mistral's Vibe coding agent got an official macOS and Windows app that runs local and cloud agents side by side. Anthropic's linked billing beta lets customers pay for Claude inside third-party platforms such as Linear and Ramp without handing over API keys.
Devtools & Infra
- AI security scanning that hands maintainers fixes: in a pilot on six open source projects (including Caddy, Express, ImageMagick and Netty), Microsoft's scanning system flagged thousands of possible issues; after AI triage and hand review by seven researchers, maintainers received 99 private reports, each with a proof of concept and a suggested fix. They had fixed 55 by November 15. GitHub concedes the checking does not scale yet: it took the researchers about two days per project. Why it matters: maintainers are flooded with unverified AI findings, and this puts the cost of verification on the sender.
- Agent sessions that outlive the window: GitHub described the Agent Host Protocol, an open standard built first in VS Code that lets a Copilot agent session started in one app continue in another, since an agent tied to an editor window dies when it closes. Why it matters: agents that work for hours need to live somewhere more durable than the app that launched them.
- Sandboxes that sit outside the agent: Simon Willison and Prime Radiant released three "external sandbox" servers (WebAssembly, Apple Containers and QEMU versions) sharing one local API: the coding agent stays outside and only the code it runs goes in. An optional proxy limits which hosts that code can reach and swaps in credentials so the agent never sees the secret. Why it matters: it makes trying a local model safer, since a mistake or a prompt injection runs inside the box.
- curl's maintainer on agent-made Rust rewrites: Daniel Stenberg counted agents turning curl's 185,000 lines of C into 340,000 lines of Rust in a little over three weeks; in the same period the real project merged 300+ commits, about 200 of them bug fixes the clones do not have, Simon Willison quotes. Why it matters: a rewrite copies one moment of a codebase, and the cost of software is keeping up with it afterwards.
Security & Privacy
- Sources: the outages came from a rush to patch: Wall Street executives and the NSA told members of Congress that this week's failures stemmed from systems overwhelmed by a rush to patch vulnerabilities found by AI models, the Washington Sun reports. Secretary of State Marco Rubio said there is "zero evidence" the FAA, Microsoft and bank failures are connected and that the US is not under systemic attack. The FAA paused installations of Juniper routers like the one officials believe contributed to Friday's ground stop (owner HPE says the device was not defective), and Apple became the latest to stumble, blaming an "internal update" for an outage of some services. Why it matters: yesterday the Homeland Security secretary blamed new technology on old systems; this account is more specific, and it says the damage came from the defense, not from an attack.
- Nginx: a second AI-found server flaw in two days: security firm Xint published NGINX2Root, remote code execution vulnerabilities in the web server that sits in front of a large share of websites, and told Forbes it found them using OpenAI's GPT-6.1 Sol and GPT-6 Luna. Wiz says it found the same flaws separately and describes them as reachable before login, in Nginx's HTTP/2 handling. Why it matters: two teams reaching the same bugs at once suggests these finds are now a matter of who runs the models first, and it hands operators another urgent patch a day after Java4Shell.
- Java4Shell: remote code execution in Java, found with an AI model: security startup Calif disclosed a flaw in OpenJDK's Base64 processing (CVE-2026-114938) that lets an attacker run code on a server without logging in; it says it used OpenAI's GPT-6 Cyber to find it, per the New York Times. Oracle has released a patch. Cloudflare's write-up rates it 9.2 out of 10, says Calif showed attacks without a login against default installs of Keycloak, Jira, Confluence, Artifactory and Jenkins, and ships firewall rules in log-only mode, with the warning that only updating Java fully fixes it. Why it matters: Java runs a large share of server software, so the fix has to reach many systems quickly, in a week when hurried patching is itself being blamed for outages.
- Optimizely confirms the breach as the payload comes into focus: Optimizely told Brian Krebs an attacker exploited a vulnerability in its experimentation product, since fixed, to add malicious JavaScript to customer projects. The code was rebuilt 26 times in a day; it stole card data, cookies and screenshots, swapped cryptocurrency addresses, and served exploit kits aimed at older iPhones and iPads. Only 24 affected sites had a content security policy (a browser rule limiting where scripts load from) that fully blocked it. Why it matters: yesterday's entry had the reach; this is the mechanism, and a defense most sites had not switched on.
- Microsoft's agent-framework package compromised on PyPI: Socket reports the official
agent-frameworkPython package was published with a credential stealer. Why it matters: it is the library developers install to build agents on Microsoft's stack, so a poisoned release lands on machines that hold cloud and API keys. - Quad9 pulls its DNS resolver out of France: after a French court imposed daily fines on the Swiss nonprofit for not blocking domains tied to alleged copyright infringement, Quad9 is withdrawing its free resolver (the service that turns site names into addresses) from the country. Cloudflare warns that Quad9 cannot notify users, among them hospitals and government offices, whose connections will simply fail, and that a Belgian order threatens Cloudflare, Google and Cisco with penalties from early January 2027; the three have appealed. Why it matters: copyright blocking orders are now removing basic internet infrastructure from whole countries.
- Instagram's school safety program doubled as teen recruitment: internal Meta documents from child safety lawsuits, reported by NPR, describe "Project Gryffindor", a school program meant to "fuel organic and unattributed acquisition" of users; a November 2023 presentation says it would give teens a new reason to join Instagram. Why it matters: the School Partnership Program was presented to schools as a safety measure, and the documents show growth goals behind it.
- Cities and voters drop license plate cameras: San Diego's council voted to end its contracts with Flock and Ubicquia, one of the largest cities to cut ties with Flock. The EFF counts ballot wins against plate readers in York, Maine, where nearly three-quarters of voters backed removal, and in League City, Texas. Why it matters: opposition to these cameras has moved from council comment periods to votes that cancel contracts.
Startups & Industry
- Nvidia posts a record quarter and tells Huang no: revenue was $115.8B, up 103% from a year ago and above the $109.1B estimate, with data center at $107.6B. Chair Harvey Jones opened the call by saying the board unanimously accepted Huang's resignation and "does not intend to reinstate Jensen as CEO", then refused questions on it. Semafor's sources say a faction of directors is still pushing for his return. Why it matters: the board's position was a source's word until now; it is on the record, and the numbers gave directors no financial reason to reverse.
- Nvidia ends H200 sales in China: interim CEO Colette Kress said Nvidia is "no longer shipping" its licensed H200 chips there and has none in its guidance, without saying why. That ends its part in the deal Huang and President Trump brokered, which let the US collect a 25% cut. Why it matters: the first strategic decision after Huang undoes the China arrangement he negotiated personally.
- The Nvidia inquiries reach the SEC and Congress: a civil SEC probe is examining whether the company's disclosures misled investors, Wall Street Journal sources say. House Democrats on the Judiciary and Foreign Affairs Committees opened an investigation into Commerce Secretary Howard Lutnick and his family's ties to Nvidia, which Politico reads as groundwork for oversight in 2027, and NBC News reports national security and AI figures are urging the Justice Department to examine what Nvidia knew about chip smuggling. Why it matters: yesterday the Justice and Commerce Departments confirmed their inquiries; now a securities regulator and Congress are asking too, and one of the questions is about the official who oversees export rules.
- Anthropic weighs a bigger IPO at a higher price: CNBC's sources say the company is considering upsizing the offering and pricing it well above $80 a share, past the $70 to $78 range, as demand surges. The Center for AI Safety and others answered with the "MisAnthropic Statement", challenging the company's safety branding and what it calls a reckless attempt to build AI that improves itself. Why it matters: yesterday investors were guided to the top of the range or higher; today the range itself looks too low, and the safety pitch that helped sell the company is being contested by safety researchers.
- Washington and Chinese AI, both directions: Reuters sources say the Trump administration is considering making US companies disclose on their websites when they use Chinese AI models, to discourage it. ByteDance is in advanced talks to spend about $500M more on Microsoft's Azure, mainly to reach OpenAI models, which has drawn complaints from OpenAI, per The Information. Tencent president Martin Lau said the new US rule on remote access to chips "sets everyone back in China" and will delay model training. Why it matters: the US is trying to keep Chinese models out of American products while a Chinese giant buys American models through a cloud reseller, and the remote access rule covered Monday is already biting.
- California bans Waymo from freeways: the state barred the robotaxis from its freeways indefinitely after a viral video showed one failing to yield to the highway patrol, and Waymo halted freeway service nationwide. Ars Technica reports the DMV's order, which stands until Waymo shows its cars no longer "pose an imminent hazard to public safety", centers on robotaxis driving into construction zones at speed, the same problem behind a June recall of nearly 3,800 vehicles. Why it matters: one incident on video removed a capability across every city Waymo serves, not only in the state that issued the order.
- Deals: Ramp is in early talks to raise $1B or more at a $70B+ valuation, its third round in a year, per the Wall Street Journal. Semafor's sources say Apollo's $40B chip financing for SpaceX runs through a special-purpose vehicle, so SpaceX rents the GPUs and keeps the debt off its balance sheet. Why it matters: the largest AI hardware purchases are increasingly being made by entities built to hold the debt somewhere else.
Research
- Telling a coding agent its budget does not help: in 5,835 runs on two program-repair benchmarks, enforcing a token cap cut tasks solved by 20 percentage points while using about a twelfth of the tokens per solved task; disclosing the cap to the agent never improved results, in any wording. Why it matters: cost control for agents is a matter of how the limit is enforced, not of what the prompt says about it.
Elsewhere
- Vulcan failure traced to fuel supply, not the engines: ULA's initial review of Sunday's flight data says the low booster thrust came from inadequate fuel supply from the rocket to its BE-4 engines, not from how the engines performed, though the root cause is still open. Why it matters: Monday's digest had the loss of 40 Amazon satellites; this moves the fault from Blue Origin's engines to ULA's own vehicle, which decides who has to fix what before Vulcan flies again.
- Voyager 1 is one light-day away: at 5:16AM ET on Wednesday the probe passed 16.094 billion miles from Earth, the distance light covers in 24 hours, 49 years after launch. Why it matters: it is the farthest any human-made object has travelled, and a radio signal now takes a full day to reach it.
- Also: GTA VI launched to a 99 on Metacritic and 98 on OpenCritic, records for both, and Take-Two shares rose more than 6%, Bloomberg reports.
Hacker News
Code hosting wobbled: Microsoft's ONNX Runtime repository disappeared from GitHub (217 points) and Codeberg was reported unreachable, a reminder of how much tooling hangs off two or three hosts. The longest argument was over what AI answers do to the sites they draw on: AI is killing All About Berlin, one site owner's account, ran to 346 comments (discussion). A write-up of Harvard research finding that coding agents moved the bottleneck to code review drew more comments than points (172 on 120).
The top of the page belonged to stories covered above: the GPT-6.2 launch at 985 points (discussion), the Apple outage at 581 (discussion), Quad9's exit from France with two separate front-page posts, and by evening the Nginx flaw at 314 beside Java4Shell.
Threads
- Found by AI, patched in a hurry: the account given to Congress blames the week's outages on patching AI-found vulnerabilities; Java4Shell and the Nginx flaws are exactly such finds, a day apart, and the GitHub and Microsoft pilot shows the same supply of findings reaching open source maintainers with fixes attached.
- Nvidia after Huang: the board's refusal, the end of H200 sales in China, the SEC probe and the House inquiry into Lutnick arrived with a record quarter; the company is unwinding his China deal while a faction of directors still wants him back.
- Chips and China from both sides: Nvidia stops shipping to China, Tencent says the remote access rule sets it back, ByteDance turns to Azure for OpenAI models, and Washington weighs labelling Chinese models inside US products.
- Price as the weapon: OpenAI cut Astra to match Claude Fable 5.2, and Microsoft pitched MAI Code 2 at a fraction of both, while Epoch's workplace test and the arXiv coding paper ask what the extra benchmark points buy.
- Safety talk meets the money: Christiano's call for a pause and the MisAnthropic Statement landed on the day Anthropic's IPO price moved up and chip earnings showed no slowdown; the persona-training paper is a concrete case of the monitoring those warnings depend on getting weaker.
- What agents cost after the demo: Stenberg's count of bug fixes the Rust clones lack, the budget experiment showing enforcement beats instructions, and the Agent Host Protocol and external sandboxes all deal with agents that run long and need supervising.