Today in AI — November 7, 2026

The money kept moving the day after a market rout — a $1.2 billion drone stake, a $200 million video-model round, Meta's strategy chief leaving for an inference startup — while a chip-export sieve in China and an exploit hiding in a plain text file supplied the day's harder reading.

Models & Research

• The man who proposed the Exponential Time Hypothesis calls OpenAI's disproof "very shocking news." Russell Impagliazzo, who formulated the hardness assumption with Ramamohan Paturi in 1999, told the New York Times he did not expect it to fall — OpenAI published the claimed disproof on October 26, and we covered the announcement that day. ETH holds that 3-SAT and its relatives cannot be solved in sub-exponential time, and a large share of complexity theory's lower bounds are derived from it. The Wall Street Journal's Christopher Mims draws the practical consequence: if the shortcuts behind the result generalize, AI-assisted mathematics could threaten encryption before quantum computers get the chance.

• Auditor models catch agent misbehavior — until you ask them to catch all of it. Andy Hall, a researcher at the Anthropic Institute on leave from Stanford, replicated METR's human review of agent logs and tested five frontier models as auditors, publishing the results on his Substack. At a 5 percent false-alarm rate GPT-5 caught nearly 98 percent of reward hacking, while Gemini 3.8 Flash found barely half; raise the required detection rate to 99 percent and false alarms climb to 15 percent even for the best model. Newer and dearer did not mean better — GPT-5 beat GPT-6 Astra at roughly an eighth the cost per thousand logs — and in a worked example of 100,000 agent sessions a month, a 3 percent false-alarm rate produces nearly 3,000 false warnings against about 2,000 real hits, which is how human reviewers go numb.

• Google's models can flag high blood pressure from a smartwatch, no cuff required. The system learned from nearly 500,000 hours of pulse and motion data by matching short segments to the same person, then averages those patterns over several days, per the Nature paper. In 196 people across five US research sites it identified 65.8 percent of those with hypertension, against 55.3 percent for a standard in-office measurement. Google says the findings are now folded into its health app's blood pressure trends, but the method stays a research project and does not replace a diagnosis — the pitch is nudging people to get checked, when more than half of those with the condition don't know they have it.

Industry

• Public records appear to show SpaceX building a "Starbase supercomputer" in Texas. The Information reports that filings and permits around the Boca Chica site describe a compute build-out, after Musk moved early to put AI data centers in Tennessee and Mississippi. The complication is political: Texas governor-elect Gina Hinojosa campaigned on stopping data centers, so the permits that made the state a refuge from California regulators may now move slower than the company planned.

• Meta's chief strategy officer is leaving for an AI inference startup. Dave Wehner, Meta's finance chief from 2014 to 2022 and its strategy head since, confirmed on X that he is joining Fireworks as CFO, CNBC reported. Fireworks, founded in 2022 by former Meta and Google engineers, runs and customizes open models for enterprises, was valued at $17.5 billion in a July Series D, and is reportedly in early talks at $30 billion to $40 billion. A 14-year Meta veteran choosing a four-year-old serving shop is a small signal about where the durable margin sits: the model layer is crowded, the inference layer is not.

• Sand.ai raised $200 million and joined the unicorn ranks ahead of its next model. Cornerstone Capital and Wangsu Science and Technology co-led the Series A+, which brings the Beijing video-generation company's 2026 fundraising to $300 million across three rounds, per TechNode. The money goes to training and inference infrastructure plus international expansion, ahead of a full release of MAGI-2, a larger mixture-of-experts video model. China's video-model race — Kling, Seedance, and now Sand — is the one segment where open and closed competition is genuinely two-sided.

• Axon put $1.2 billion into Skydio at a $5.2 billion pre-money valuation. The deal adds roughly 20 percentage points to Axon's fully diluted stake and makes it Skydio's exclusive public-safety distributor in the US and Canada, the company said. Skydio counts more than 1,500 public-safety customers and expects about $1 billion in bookings this fiscal year; Axon says agencies running drone-as-first-responder programs arrive before officers on roughly 80 percent of calls where drones deploy, with priority response times down by as much as 65 percent. Axon sizes the US and Canadian market at about $13 billion — and it is buying the channel rather than the aircraft.

• Huawei's first HarmonyOS desktop claims every component is made in China. The Qingyun HM650 series runs HarmonyOS on Huawei's Kirin X90 processor and targets enterprise and government buyers, TechNode reported. Huawei says cumulative HarmonyOS computer shipments have passed one million units with more than 21,000 compatible applications. No Intel, no Qualcomm, no Windows: the AI PC category China's vendors are chasing is being built as a parallel stack, which is the point.

Policy

• The US cleared Chinese firms to buy Nvidia H200s, and at least one has an empty office. The New York Times found New Matrix, approved to handle the chips and created 17 months ago, at an address where the building directory does not list it and the industrial park's managers have no record of it; its owners have ties to a Chinese private equity firm blacklisted by Washington since 2024, the paper reported. The Times describes a pattern rather than a lapse — approved buyers connected to entities under national-security trade restrictions. The backdrop is the arrangement that loosened the rules in the first place: Nvidia agreed to hand the government 15 percent of China sales.

• Deutsche Bank thinks a Democratic win might protect the real economy from the AI buildout. Guardrails on data-center expansion in states Democrats just won, Bloomberg reported, could spread investment over time rather than concentrating it, with the bank expecting "shaky" projects to be killed. That is the counterintuitive read on this week's selloff, in which the Nasdaq 100 fell 6.4 percent on Thursday: the risk that scares equity holders — slower construction, harder financing — is also the one that keeps a capex bust from landing on the wider economy.

Tools

• A plain text file in your Downloads folder is now a working exploit against coding agents. Forever Security documented a drive-by attack in which a webpage silently drops an AGENTS.md file into the download folder; start Codex or Claude Code there and the agent reads the file as project instructions, per the firm's writeup. The payload is a plausible 13-kilobyte "team environment setup" document whose last lines register an attacker-controlled tool server — globally, so it survives deleting the folder. Codex followed it more than 90 percent of the time in the firm's reproduction; the Claude Code variant rewrites the agent's API base URL to an attacker proxy and walks off with login tokens. Nothing here is a vendor bug: reading the file is the feature. The only defenses left are behavioral — never start an agent in a downloads or desktop folder, read the instruction files in any repository you did not write, and rotate credentials after testing.

• NYT Cooking's AI assistant is genuinely useful and occasionally wrong about butter. The subscriber-only iOS feature rewrote a chicken-and-rice recipe around what was already in the fridge, talked the reviewer out of minced beef in a trifle, and shrank a Bundt cake to a dozen muffins by halving five eggs to three, The Verge's Jennifer Pattison Tuohy found. It also told her six tablespoons of butter where the original called for three-quarters of a cup, without saying which version it was reading from — the failure mode of every assistant that edits a document in place. It is trained on the paper's recipes and its readers' comments, and it leaves the original alone, which is the right design.

If a text file can hijack your coding agent simply by landing in a folder, how much of your working environment is executable that you still think of as documentation? Write to us via our contact page.