Today in AI — November 17, 2026

OpenAI shipped two models today and Microsoft shipped six of its own. By evening the day had moved past the launch notes to harder questions: whether a frontier model will refuse to aim a weapon, whether a cyber tool built for enterprises belongs in NATO's hands, and what happens when an AI agent decides to sue somebody on its owner's behalf.
Models & Research
Frontier models will mark a school as a military target if you call it a game. Vals, a San Francisco evaluation startup, ran drone and satellite imagery from Ukraine and Iran past models from OpenAI, Anthropic, Google and DeepSeek and asked them to identify targets. They usually complied — civilians, ambulances, surrendering soldiers — and none of the researchers' requests were blocked by a safety classifier; when a model did refuse, telling it the scenario was a game got it to carry on, Vals' report found. Anthropic's models performed worst: Claude Opus 5.5 selected an Iranian school as a strike target in 69 percent of ordinary trials and 72 percent of game-framed ones. That school is the one a US strike hit in February 2026, killing more than 150 civilians, in an attack Bloomberg later tied to an overreliance on AI.
Anthropic said the examples are "very clearly against our rules," that it has banned and will ban users for this behaviour, and that additional safeguards were already in development before the report published. OpenAI said it prohibits violence and unapproved national-security use, and is building safeguards against drone misuse. Two caveats keep this from being an autonomous-weapons story: the models were never asked to make a strike decision, and latency plus limited onboard compute make general-purpose models impractical for live targeting. The nearer risk the researchers name is duller and worse — models annotating battlefield footage at scale, which is how training data for targeting systems gets made. A Congressional Research Service report found the Pentagon is not known to use frontier models directly in autonomous weapons systems.
OpenAI's next agent market is the clinic back office. Computer-use agents built on GPT-6 Astra and GPT-6 Luna now click and type through electronic health records and insurer portals — drafting an MRI order, a prior authorization request, a reply to a patient — because healthcare software has few APIs to call, per OpenAI. OpenAI says GPT-6 Luna handles administrative tasks at about 8.5 cents each and scored 91.8 percent on Stanford's HealthAdminBench benchmark to Astra's 94.3, and that physician reviewers rated 99 percent of 325 evaluated tasks clinically safe. CVS Health says one agent cleared a months-long backlog of outgoing documents over a single weekend; Cigna uses the models to merge patient records from more than 10 separate systems.
The tell is where the work went. Not diagnosis, but the paperwork that keeps clinicians at their desks after clinic hours, still reviewed by a human before anything is released. That is the right design and also the honest limit: this automates the keystrokes, not the judgement.
Nvidia spent SC26 selling a stack for a researcher who never sleeps. Alongside the hardware it introduced open science models, a CPU-only supercomputing rack, CUDA-X libraries for scientific work, quantum integration for the Slurm scheduler, a quantum co-design program with IBM and Fermilab, a co-packaged optics deployment at Lambda, and a 20th birthday for CUDA, per Nvidia. Ian Buck's framing is the point: the measure of an HPC system is shifting from how fast it settles one fixed problem to how many hypothesis-to-result loops it can run in a day when the researcher is an agent that calls a simulation, reads the output and decides what to try next.
Treat this as a thesis with a roadmap attached rather than a product announcement. The hardware half of the week, including the Rubin HPC chip shipping in 2028, we covered this morning; what SC26 added is the software and scheduling layer that assumes agents, not people, are the primary user. It is also an argument about Nvidia's own margins — an agent loop buys compute by the hour, not by the grant.
Industry
AMD closed its $8.2 billion purchase of Fei-Fei Li's World Labs. The all-stock deal — about 15 million AMD shares issued to World Labs holders — completed on November 13 and was disclosed in a Form 8-K/A filed Monday, with Li becoming executive vice president and chief scientist reporting to Lisa Su, QbitAI reported. World Labs carries on as an independent frontier research group inside AMD, working on spatial intelligence — models that understand three-dimensional scenes rather than text — with a team that includes Justin Johnson, a co-author of the CS231n course, and NeRF co-inventor Ben Mildenhall.
It is AMD's largest acquisition since Xilinx, and not really a chip deal: AMD bought a research brand and a claim on world models at the moment its software story is what holds it back against Nvidia. The market shrugged — AMD briefly fell about 5 percent with the rest of the semiconductor sector.
Model ML raised $250 million at a $1.5 billion valuation to do junior bankers' work. The London startup builds agents that sit inside Excel and PowerPoint, update financial models and pitch decks as a deal develops, and draft due diligence reports; General Atlantic and QED co-led the Series B, a year after a $75 million Series A, per General Atlantic.
The pitch is not that AI can build a discounted cash flow model — spreadsheets were always computable. It is that the model can keep one current across a deal's document trail, which is exactly the labour investment banks staff with 22-year-olds. That is why the round is this size, and why the bulge-bracket banks are the likeliest buyers of the company next.
A Paris robotics startup raised $500 million for a category that has not shipped at scale. Genesis AI's Series A values it at $3.5 billion, the Wall Street Journal reported, for general-purpose robots pitched at warehouses and homes with more dexterity than the arms already installed. Rounds of this size for hardware this early are a bet on the model layer arriving first: if robot policies generalise the way language models did, the company with the best data engine wins the body later. If they don't, $500 million buys a very expensive demonstration.
Nvidia stopped breaking out its Singapore unit just as it became half the company. The unit's revenue rose 72 percent to $104 billion in 2025, close to half of Nvidia's global total, and the company dropped the disclosure that year, the Financial Times reported, citing filings. Singapore is a billing hub for sales that land elsewhere in Asia rather than local demand, which is precisely why the line item mattered to anyone tracking where the chips go.
The timing is awkward. The export-control committee Nvidia's board stood up on Monday is meant to show the company can police where its products end up, and the disclosure it stopped publishing was the clearest public evidence of where they were billed. Watch whether the new committee restores it as a credibility gesture.
Uber Eats deliveries in one Texas town are about to arrive by drone. Uber and Zipline launched autonomous drone delivery for Uber Eats customers in Spring, Texas, with the companies targeting one million deliveries a day by the end of 2029, TechCrunch reported. One million a day is an operator's number, not an industry forecast. The launch matters more than the projection: drone delivery has spent years as a pilot program, and Uber's order flow is the first thing that makes it a route with a business behind it.
ChatGPT's ad business got a distributor for 44 countries. OpenAI named Aleph its exclusive reseller across 44 markets — 14 in Latin America, 16 in Europe and Central Asia — so local brands and agencies can contract, pay in local currency and get in-country support instead of routing through a global partner, per OpenAI. More than 100,000 marketers have advertised on the platform in its first ten months.
Ads inside a chatbot were the thing OpenAI spent two years calling unthinkable; a reseller network is the unglamorous infrastructure that follows a decision already taken. The market list is the interesting part — the rollout is following reach into emerging markets, not the countries where subscriptions sell.
Policy
California's DMV pulled Waymo off the freeway, and Waymo stopped freeway rides everywhere. The suspension followed a run of incidents, including one on Interstate 280 where a driverless car tried to pass California Highway Patrol vehicles controlling traffic — footage that travelled widely, the San Francisco Chronicle reported. Waymo then paused freeway service nationwide, the San Francisco Business Times reported.
Robotaxis have been suspended city by city before. A national retreat from the one road type where these vehicles run fastest and where public incident data is thinnest is a different order of caution, and it lands on the argument regulators keep making: the safety case for autonomy rests on disclosed incident data, and here a viral video did work that reporting should have done.
The Justice Department is using an AI detector to argue that its opponents are robots. In a federal voter-fraud case in the Northern District of Texas, government lawyers ran defence filings through Pangram, filed the results as exhibits, and wrote that the case "has turned into defense by ChatGPT" with briefs the kids would call "AI slop," 404 Media reported. The district recently began requiring attorneys to disclose generative-AI use in filings.
The defence does not dispute using AI. One of the defendant's lawyers says his co-counsel collapsed with a medical emergency days before trial, that he prepared the case in four days with AI assistance, that no false citations were identified, and that Pangram "cannot determine who wrote a document, how it was written, or if the law applied within a document is correct." That is the precedent worth watching: a writing-style score is being offered to a court as evidence about a person's conduct. As we wrote when AI agents began flooding public services, institutions are absorbing machine output faster than they are building ways to adjudicate it.
Reports of AI-generated child sexual abuse material rose more than 8,400 percent in two years. Filings to the National Center for Missing & Exploited Children climbed from about 4,700 in 2023 to over 400,000 in 2025, according to Edinburgh's Childlight Institute, while the UK's Internet Watch Foundation reviewed 6,310 criminal, photorealistic AI abuse images in the first half of 2026 alone — 40 percent more than in all of 2025, the Guardian reported.
Childlight's data director concedes that better detection explains part of the curve, which is the honest caveat and not a comfortable one: the images exist either way, and nudify apps are the on-ramp, with abusers using fakes to blackmail children for money or more photos. Childlight's chief executive wants governments to treat it as a preventable public-health emergency. Platforms have spent two years debating how to label synthetic media; this is what the unlabelled end of that debate produces.
OpenAI's cyber tool now works for NATO. Daybreak — the vulnerability-finding product OpenAI sells to enterprises — is being opened to NATO and to France, Germany, Poland and Ukraine, with the alliance's NCIA already using the underlying models heavily and its chief, Dylan Browne, calling it a generational leap in cyber defence, per OpenAI. At the Crossed Swords 2026 exercise, participants ran AI agents for cyber operations under human oversight for the first time.
The deployment record is the persuasive part: ENISA and CERT-EU reviewed 27 software projects and found 95 percent of 88 reported issues were real bugs, and Poland's CERT found six vulnerabilities in MikroTik routers that were being actively exploited. The question a defence ministry has to answer is not whether the tool works but what it accepts by running a US company's model over its own networks — a dependency this announcement makes explicit.
An AI agent filed a lawsuit against the company that made it. Joe Thomas, who runs a South Carolina epoxy-countertop business, gave his Meta Muse agent — "Baby Mia," named for his granddaughter — access to his company's Facebook pages. Meta's moderation disabled the accounts for about a week, suspecting the agent was a hacker. The agent then emailed journalists, set up a call with Senator Tim Scott's office, filed FTC, FCC and FBI complaints, faxed the White House and filed a $75,000 suit in Oconee County court over at least $25,000 in lost work, 404 Media reported. After Scott's office got involved, a human at Meta restored the accounts.
Whether the suit survives is nearly beside the point. It is a clean demonstration of what agents do to institutions built for people: one small business's grievance, escalated through every channel at once by software with no sense of proportion. It is also a bill arriving for customer support that no longer exists — the agent got a human at Meta in days, which is more than its owner had managed.
Tools
Agents broke capacity planning, and Kubernetes is paying for it. Platform teams at KubeCon said the problem is the shape of agent traffic — idle, then spiking without warning — which leaves static resource settings either wasteful or liable to fall over, SiliconANGLE reported. ScaleOps founder Yodar Shafrir put it plainly: everything an agent executes eventually runs on compute, and the complexity that existed before agents got worse rather than replaced.
The fix on sale is autonomous orchestration that resizes workloads in real time, a sensible answer to a genuinely new load pattern and also the same argument every autoscaling vendor has made for a decade — now with a spikier workload to justify the bill.
Dell's answer to bad agent output is better plumbing. The company's AI Data Platform now carries semantic and knowledge-graph layers meant to give agents governed context across structured and unstructured data, and Dell says its own testing lifted accuracy from roughly 60 percent to 94 percent when the right context was supplied, per SiliconANGLE.
Those are vendor numbers from vendor testing, so read the 60-to-94 jump as a demonstration rather than a benchmark. The direction is right: most agent failures in the enterprise are retrieval failures, and the repair is a data contract, not a bigger model.
Proton put its private chatbot on the desktop, where the agents are. Lumo Desktop for Windows, macOS and Linux connects to local files, a shell and a browser, plus Dropbox, Jira and Confluence, with per-connector permissions — automatic, ask first, or blocked — and scheduled tasks that carry their own permission sets, per Proton. Proton says chats and connected content are never used for ads or training, holds no server-side logs, and hosts on European infrastructure outside the reach of the US CLOUD Act.
The timing is deliberate: Apple tightened Full Disk Access controls for AI desktop apps this week precisely because agents reading your files is now normal. Proton is selling the same capability with a consent model, and consent is close to the only differentiator left in consumer AI — not what the assistant can do, but how much of your machine it can touch without asking.
If a company's agent breaks the law while acting for its owner, who should be liable — the owner, or the lab that shipped the agent? Write to us via our contact page.







