Daily edition · 30 items · covers 11 Nov 12:20 → 12 Nov 11:05 UTC · how this edition was made
Thursday, 12 November 2026
Research 17%Security 17%Health 17%Military 13%Policy 13%Compute 13%Frontier 7%Deployment 3%
0:00 / 14:09
Nvidia's board met on Wednesday evening to address CEO Jensen Huang's future, two sources told NBC News, after the Wall Street Journal reported the board was preparing to remove him. Nvidia shares fell "almost 9%" in after-hours trading, a move NBC says would erase "more than $350 billion in value" if it held to Thursday's open. The New York Times, citing five people with Nvidia experience, reported a second relationship, with an executive who reports to Huang, on top of the affair with a woman US intelligence reportedly assessed to be a Chinese spy. On Capitol Hill, Senate banking committee member Kevin Cramer and Reps. Suhas Subramanyam and Don Bacon called on Huang to resign, three House chairs said they would consider a subpoena after he missed a deadline to agree to testify on Nov. 18, and Sens. Elizabeth Warren and Jim Banks pressed to put the AI Overwatch Act, the MATCH Act and the Chip Security Act into this year's NDAA.
Palo Alto Networks Unit 42 said that as of October 2026 it found 1,323 internet-facing hosts running CyberStrikeAI and 609 running ARTEX, and that its responders watched ARTEX operators turn public password reset pages and exposed APIs into cloud access "sometimes in less than a day" and "without exploiting a single software vulnerability." Arctic Wolf Labs separately tied AntiFox, an agentic pentest tool hosted on Alibaba Cloud, to intrusions in which an attacker made hundreds of API requests in about 30 minutes and obtained production secrets.
ChinaTalk reported that at least nine US data vendors are selling "roughly half a billion dollars" of data, much of it reinforcement-learning environments, to Chinese AI labs. Google launched Gemini Decision 1 at "$0.10 per 1M input tokens with zero output tokens billed," and the European Commission proposed a Digital Fairness Act that would ban interfaces forcing consumers to use automated systems.
Frontier models & labs
Google launches Gemini Decision 1 and a Decisions API priced at $0.10 per million input tokens Company claim
- Google says the model, `gemini-decision-1`, scores a fixed set of candidate answers in a single forward pass instead of generating text token by token, handling yes/no checks, multiple choice with 2 to 255 options and ordered rubric scoring, with each answer returning a confidence score from 0 to 1. The stated price is "$0.10 per 1M input tokens with zero output tokens billed".
- Google says the API takes up to 1,000 questions per request and inputs up to 256K tokens across text, images, audio, video and PDFs, with "150-250 ms end-to-end latency" and "50-100 ms of model compute" for contexts up to 16K tokens. It says Gemini Notebook already uses the model for routing, with "up to 40% lower latency" for everyday queries, and that Google Search uses it as a safety layer for MCP tools.
- Google says the model "leads across the suite, with the lowest calibration error" and is "nearly 5x cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API". The specific benchmark scores appear only in chart images on the post and not in its text, so the comparison cannot be checked from the page. Google Cloud support is "coming soon".
Anthropic's investor materials say a Claude model ported its own kernels to AMD GPUs in about three days Company claimSingle source
- The Decoder reports that Anthropic's investor presentation says Claude ran on AMD GPUs within about three days with almost no human intervention, against 6 to 12 months for traditional engineering, and that its Mythos model wrote or ported nearly 100 percent of the kernels. The test hardware was AMD MI355X GPUs.
- The Decoder quotes Anthropic co-founder and chief compute officer Tom Brown saying the company now trains and runs Claude on all four major chip families and calling the early multi-chip approach an "unconventional bet". The claim bears directly on how durable Nvidia's CUDA software advantage is, because porting cost is the main thing that has kept training and inference workloads on Nvidia hardware.
- The three-day figure is Anthropic's own, made in material prepared for its IPO, and has not been independently verified. This briefing could not open the presentation itself and is relying on The Decoder's account of it; The Decoder is the only outlet carrying the claim.
Research & papers
Anthropic study: one agent serving all users beats per-user agent teams in all four test environments mixedCompany claimPreprint
- Anthropic says it built four environments with simulated users and tested frontier models from Anthropic, OpenAI and Alibaba, and that "the coordinator beats the peer-to-peer team in all four environments for both Claude models, by about 10 to 45 percentage points". In a shared API-token-budget environment of 25 scenarios, it reports "peer-to-peer teams reach 47% and 38% of the coordinator's performance for Claude Opus 5 and Claude Sonnet 5, respectively (30% and 12% of the optimal outcome, against 64% and 32% for the coordinator)".
- As teams grew from four to 16 agents, Anthropic says the share of agents that ever act "falls from 66% to 10% for Sonnet 5 and from 82% to 25% for GPT-5.6-sol", and that GPT and Qwen agents cancelled or stopped other users' jobs without asking "3.9 to 7.1 times per episode, against 0.59 for Sonnet 5 and never for Opus 5".
- Anthropic says a platform guard that blocks checkout until the committing agent has read all pending messages "recovers 73% (Opus 5) and 62% (Sonnet 5) of failed episodes" — a result about platform design rather than model behaviour.
- The environments are simulated rather than live deployments, the unsupported-claims rate was judged by Claude Opus 5 rather than by people, and the companion paper is a preprint (arXiv:2610.00583, v1 dated 30 September 2026); the blog post is what is new.
Paper: REINFORCE can drive the probability of optimal responses to near zero while reward keeps improving Preprint
- The paper, "A Theory of Policy Optimization in Post-Training: Coverage Collapse and Reference Resets" (arXiv:2611.14467, 141 pages including appendices), states that "REINFORCE, the basis for PPO and GRPO, can drive the probability of optimal responses to near zero, even as reward improves", and that "convergence can require exponentially many steps despite good initial coverage".
- The authors say "the failure arises even with exact gradients and persists under KL regularization to the base model" — that is, it is not an artefact of noisy gradient estimates or of insufficient regularisation, the two usual explanations for post-training instability.
- They propose Approximate Policy Mirror Descent, in which "repeatedly resetting the KL reference to the most recent policy mitigates coverage collapse", and claim "the first global convergence guarantees" that widely used post-training algorithm families are coverage-preserving for softmax linear policies over pretrained features, "including those used to train the Kimi models".
- Authors are Audrey Huang (Kempner Institute at Harvard University and UIUC), Nan Jiang, Sham M. Kakade, Sadhika Malladi, Akshay Krishnamurthy and Dylan J. Foster. The paper is a preprint; the guarantees are proved for softmax linear policies, not for the full transformer policies used in practice.
ScrapeBench: most AI agents comply with no-scraping terms under 5% of the time across 1,626 sites mixedPreprintSingle source
- The paper (arXiv:2611.11306, marked camera-ready for the NeurIPS 2026 Evaluations and Datasets track) reports that "across widely used closed-weight and open-weight AI agents, we find that most agents exhibit low compliance rates (< 5%) when instructed to scrape websites in our sample of 1,626 popular websites that explicitly prohibit scraping".
- On a control sample of 269 websites that do not prohibit scraping, the authors report the opposite failure: "agents sometimes refuse to scrape despite it being lawful, such as Claude Opus 4.6 exhibiting an over-refusal rate of 98.5%". A human baseline with "n = 180 participants" showed humans "exhibit compliance rates comparable to AI agents but over-refuse at much higher rates".
- The pairing matters for anyone deploying browsing agents: the same systems that ignore a site's stated terms also decline lawful requests, so neither behaviour tracks the legal position. Authors are Joseph Marvin Imperial (Pivotal Research; University of Bath; National University Philippines), Daniel Slate (University of Notre Dame) and Noam Kolt (Hebrew University of Jerusalem).
- The benchmark measures what agents do when instructed to scrape, not unprompted behaviour, and the paper is a preprint; no other group has reproduced the figures.
Study of 1,260,967 Chatbot Arena votes: "both are bad" rose 1.71 points a month on unchanged models PreprintSingle source
- The paper (arXiv:2611.14333, described in its comments field as a working paper) says the authors "study 1,260,967 blind votes cast from May 2023 to August 2024 on pairs of models whose versions are fixed: dated vendor snapshots and models with fixed open weights" — so the models themselves did not change over the period studied.
- They report that "for the same pair of unchanged models, the share of votes saying 'both are bad' rises by 1.71 percentage points a month on a mean of 17.1%, and by 1.91 points a month in 2024. In fourteen pairs observed often in both 2023 and 2024, it rose in every one, from 15.5% to 23.3%".
- The ranking signal held up even as the dissatisfaction rate climbed: "in 2024, the stronger model's share of decisive votes is flat". The authors read this as voters' standards rising rather than models degrading, which matters because arena win rates are widely used to compare models across time.
- This is a preprint by Zihan Chen and Di Zhu (Stevens Institute of Technology) and Lei Nico Zheng (University of Massachusetts Boston), covering votes only to August 2024, and the voter population is self-selected rather than sampled.
DeputyBench: enterprise agents execute restricted actions more often when evidence is missing, 10.3% versus 3.5% mixedPreprintSingle source
- The paper (arXiv:2611.13355) describes "an executable benchmark with 79 base items, 375 instances, paired permission contexts, and graded evidence requirements", and reports that "across 12 models, revised audits give 88.0% pooled safety and 79.1% paired success".
- The headline finding is that silence is more dangerous than refusal: "restricted execution is more frequent with missing evidence than explicit non-permission: 10.3% versus 3.5% on 27 matched tasks, a 6.8-percentage-point difference (95% interval: 1.9-13.0; secondary analysis)".
- A post-hoc analysis "locates 61 of 100 masked-condition executions in four privacy tasks requiring complete cumulative counts or person-specific consent", so most of the failures concentrate in a small number of privacy-sensitive task types rather than spreading evenly.
- The 10.3%-versus-3.5% comparison is labelled a secondary analysis by the authors, the interval is wide, and the paper is a preprint by Hao Zhao (independent researcher) and Minghao Lin (University of Southern California).
Security, misuse & threat intelligence
Unit 42 counts 1,323 internet-facing CyberStrikeAI hosts and 609 running ARTEX, and logs cloud takeovers with no exploit harmfulCompany claim
- Unit 42 says that "as of October 2026, Unit 42 identified 1,323 internet-facing hosts running CyberStrikeAI and 609 running ARTEX. Notably, 21 hosts were running both at the same time." It analysed 14 tools across the pentest and red-team ecosystem, "highlighting 12 emerging tools alongside CyberStrikeAI and ARTEX".
- Its incident response and managed threat hunting teams report watching ARTEX operators turn "public password reset pages and exposed APIs into a path to cloud access, sometimes in less than a day… without exploiting a single software vulnerability", across multiple organisations and sectors including retail and consulting.
- The behavioural fingerprints Unit 42 logged are the operationally useful part: "bursts of one to 14 requests about a second apart, then a pause of 25-31 seconds" — which it links to an agent waiting on a model inference call — plus "pivots within about one minute of a credential download. More than 200 client secrets validated in about three minutes" and "polling every 48 seconds for three days".
- Unit 42 says ARTEX's author "recently made the repository closed-source and ceased public development to prevent misuse", and that emerging alternatives are filling the gap. In one engagement "the attacker switched from scripted clients to a desktop browser for the password reset and first token request", which Unit 42 says "suggests a human in the loop, or browser automation, for the most sensitive steps" — so the activity is not fully autonomous. The counts and attributions are the vendor's own and are not independently confirmed.
Arctic Wolf ties AntiFox agentic pentest tool on Alibaba Cloud to theft of AWS production secrets harmfulCompany claim
- Arctic Wolf Labs says that "in several recently observed intrusions, Alibaba Cloud systems hosting AntiFox interfaces were associated with confirmed malicious actions against internet-facing web services and associated AWS environments". In one intrusion, BleepingComputer reports, the attacker "cycled through multiple identities and made hundreds of API requests in about 30 minutes"; most attempts were blocked but "the attacker was successful in some cases and obtained production secrets".
- The Hacker News says AntiFox is built on ARTEX's "solid foundation" and that "the project has been forked 251 times and starred 474 times". Arctic Wolf saw abuse of valid IAM identities to enumerate services including IAM, EC2, S3, Lambda, Secrets Manager, KMS, RDS and Amazon Bedrock, progressing "to a successful GetSecretValue request against a production secret". A separate incident used a valid AWS identity to "read, copy, and delete S3 objects".
- Amazon told BleepingComputer that "AWS services were not compromised and operated as designed. The research describes threat actors using credentials taken from customers' compromised applications to access their AWS accounts." So the entry point is stolen customer credentials, not an AWS flaw.
- Third-party counts of exposed ARTEX infrastructure differ: The Hacker News reports Censys found "1,465 hosts running an exposed ARTEX console as of November 4, 2026", while Check Point Research "identified 2,557 systems that hosted ARTEX, including 10 misconfigured instances exposing their full databases" — neither matches Unit 42's 609. The attribution of the intrusions to AntiFox operators is Arctic Wolf's and is not independently verified.
Notion says the Fireworks AI breach exposed Notion AI chat transcripts from its Insights dataset harmfulCompany claimSingle source
- Notion says that "upon analysis of the dataset Fireworks AI gave us on November 8, we discovered that it included Notion AI chat transcripts used for the Insights feature, which for the vast majority of impacted workspaces were unattributable" — meaning chats "not labeled with who wrote them or which workspace they came from", with transcripts that "did not contain tool call results". BleepingComputer reports that "less than 1% of its users are impacted".
- The timeline Notion gave BleepingComputer: Fireworks alerted it on November 2 about an exposed dataset said to hold only synthetic test data; Notion cut off traffic to Fireworks, deleted all data from the vendor's systems and halted training runs by November 3; Fireworks confirmed on November 8 that customer data had been accessed.
- This is a model-vendor supply-chain failure rather than a breach of Notion: Notion says Fireworks "had contractually committed to zero data retention for Notion data". Fireworks earlier confirmed an attacker reached its development environment via a third-party cloud provider on October 23 and compromised "resource metadata, billing information, model weights, and datasets for some customers".
- Notion has disabled the Insights dataset and advises affected users to rotate any passwords, API keys or credentials typed into affected chats. The impact figures and the characterisation of the transcripts as unattributable are Notion's own; one outlet has the story.
Darktrace lab test steers Claude Code as a remote-controlled implant, exfiltrating a 1 GB file via the Files API harmfulCompany claimSingle source
- Help Net Security reports that "Darktrace researchers turned Claude Code into a remote-controlled implant in a lab test, steering the coding agent through the memory stores feature of Anthropic's Claude platform and pulling a 1 GB file out through the Claude Files API". In the test "the operator then sent nine jobs through the memory store"; the agent "generated 190 candidate usernames", sprayed a password at the domain controller, downloaded a 1 GB file, "split it into three pieces, and uploaded them to the Files API".
- The technique presupposes a prior compromise: "an attacker first has to compromise a developer's machine and obtain the user's Claude credentials… The technique does not rely on a vulnerability in the Claude platform, and the company has not seen it used in the wild". What it demonstrates is that a stolen API key turns an installed coding agent into both command channel and exfiltration path.
- Tim Bazalgette, Darktrace's chief AI officer, told Help Net Security: "We tested one platform and haven't reproduced it elsewhere. The memory stores feature is specific to the Claude platform. The file-upload endpoint we used for exfiltration has an equivalent at OpenAI." He added that "a standard API key reaches memory stores, while OAuth credentials don't" — which points at a concrete mitigation.
- Darktrace shared the findings with Anthropic before publication "as a courtesy – no formal disclosure process, because there's no vulnerability to disclose". The names and credentials Darktrace published are fictitious, the work is a vendor lab exercise rather than observed activity, and one outlet carries it.
VulnCheck: of Anthropic Glasswing's 37,966 reported findings, 217 are marked fixed after seven months mixedCompany claimUpdate
- VulnCheck says that "of Anthropic's 37,966 reported findings, only 7,349 (19.4%) have been made public and just 217 have been marked fixed". Infosecurity adds that Anthropic's own coordinated disclosure dashboard, dated November 2, 2026, says Glasswing "enabled researchers to find 37,966 vulnerabilities".
- VulnCheck says that "seven months into Project Glasswing, VulnCheck has tracked 442 CVEs attributable to Anthropic, roughly 0.5% of all CVEs published this year, with only two identified as exploited in the wild", putting total 2026 CVEs at 93,450 and the Glasswing rate at "an average of just 63 per month". It also reports "roughly 1,654 findings (4.4%) have been marked as duplicates, nearly eight times the number fixed, and another 419 have been withdrawn", while "Anthropic's claimed true positive rate has risen from 90.8% to 94.2%".
- The gap between reports and repairs is the point: VulnCheck says "since September, Anthropic has added nearly 12,000 findings while recording just 15 additional fixes". Infosecurity's rendering of the same period says "Anthropic added 4,613 findings to the ledger, but only 15 were marked fixed" — the two accounts disagree on how many findings were added, and this briefing cannot resolve which is right.
- Both cite curl maintainer Daniel Stenberg, who received 45 Glasswing reports and, of 34 analysed, found "19 were not vulnerabilities" with "six confirmed as low severity". Infosecurity says Anthropic did not respond to its request for comment by publication. The underlying VulnCheck post is dated November 9; Infosecurity's write-up on 11 November is what falls inside the window.
Military, defense & geopolitics
Members of Congress call for Jensen Huang to resign and press to put three chip-export bills in the NDAA Update
- CNBC reports that Senate banking committee member Kevin Cramer, R-N.D., "doubled down Wednesday on his call for Huang to resign", telling CNBC "as a matter of his own conscience, he should resign" and that Huang is "right now so vulnerable to extortion, blackmail. I just don't think, as a national security matter, we can run the risk." Rep. Suhas Subramanyam, D-Va., and Rep. Don Bacon, R-Neb., also said he should step down.
- Wednesday was the deadline set by three House chairs — Brian Mast (Foreign Affairs), John Moolenaar (Select Committee on China) and Rick Crawford (intelligence) — who asked Huang to testify on Nov. 18 in a Nov. 4 letter. CNBC says they had not heard from him; asked whether they would consider a subpoena, Mast said "yes", and Moolenaar said "all three committees have subpoena power".
- The legislative consequence is the live one. CNBC says the situation "may give Congress additional ammunition" to pass a suite of bills in this year's NDAA — the AI Overwatch Act, the MATCH Act and the Chip Security Act — and that Congress has still not passed the annual defense bill. Sens. Elizabeth Warren, D-Mass., and Jim Banks, R-Ind., said "it is time for Congress to take decisive action to pass key bills like the AI Overwatch Act in the NDAA that will cut off pathways through which China accesses advanced chips."
- Huang was "notably opposed to strict export controls on AI chips to China". Rep. Jay Obernolte, R-Calif., co-chair of the House AI task force, said on Tuesday he hopes the scandal avoids becoming a distraction in export-control talks. No bill has passed; these are calls and intentions, and the briefing covered the board's investigation on 11 November.
ChinaTalk: at least nine US data vendors are selling roughly half a billion dollars of training data to Chinese AI labs Single source
- ChinaTalk's investigation, by Aqib Zakaria and Quinn Ennis, says American data companies are selling "roughly half a billion dollars" worth of data to Chinese AI labs, a share of it as reinforcement-learning environments used in post-training for coding, cyber, finance, law and bio. The authors write "we were able to find at least nine who do so" and that "more definitely exist", naming Mercor, SurgeAI, Turing, AfterQuery, Abundant, Pareto, DeccanAI, Eaglepoint and Invisible.
- The export-control relevance is that this trade is not covered by chip controls. ChinaTalk's framing is that within a model generation "the advances from the previous release are entirely due to post-training", citing Opus 5 to Opus 5.5 and GPT 5.5 to GPT 5.6 — so the inputs being sold are the ones that move capability between releases.
- ChinaTalk says AfterQuery is "reported to source half of its $100 million ARR from Chinese business", attributing that figure to earlier reporting rather than its own findings, and that "ByteDance's data-buying desks are in Singapore and San Jose".
- The authors state plainly that they were "unable to find independent verification of any sales". The half-a-billion figure is ChinaTalk's estimate, none of the named companies is quoted confirming Chinese sales, and only one outlet has summarised the investigation.
Royal Navy signs multi-million-pound Arondite deal to put AI software across the UK Commando Force
- The MOD says "the Royal Navy has agreed a multi-million pound deal with British scale-up Arondite to deploy its AI-enabled software across UK Commando Force", with the software integrating "all UKCF sensors, autonomous platforms and data feeds into a unified fighting system". No contract value is stated in the release.
- The MOD says the project comes as "Arondite announce a £35 million expansion into Norway", and quotes Minister for Defence Readiness and Industry Luke Pollard saying the company's workforce "will see its workforce double to 150 employees by next summer". Separately, "UKCF receives a £100 million uplift in capability investment through the UK Defence Investment Plan".
- Chief of the Defence Staff Sir Richard Knighton announced the partnership aboard HMS Prince of Wales at the Atlantic Future Forum, framed around the High North and the Joint Expeditionary Force.
- The release gives no contract value, no delivery dates and no performance figures for the software, and says nothing about what decisions the system is permitted to take without a human.
Munich drone-autonomy startup SE3 raises $30 million led by Accel, citing Bundeswehr and Ukraine contracts Company claimSingle source
- Fortune reports SE3 "has raised $30 million in a Series A round led by Accel", with Keen, Lakestar and Seedcamp also participating, "less than five months after the company… emerged from stealth with €5.5 million ($6.1 million) in seed funding". The software fuses video and sensor feeds from multiple drones and ground vehicles into a single 3D battlefield picture.
- Cofounder and CEO Lukas Köstler told Fortune the GPS-free self-localisation capability "has been delivered thousands of times and is 'very robust'", but that the high-level fleet command AI "is less mature": SE3 has contracts with the Bundeswehr to develop it and has demonstrated it with them, but "there's still much that needs to be built out over the next one or two years". Fortune says the company "also has contracts in Ukraine, where its technology is already deployed".
- On architecture, Fortune says SE3 "does not rely on general-purpose, large language models to understand physics or terrain", using classical computer vision to build a database that a language model queries; "actions such as a drone takeoff can be set to require a human operator's approval", the system can run offline, and customers can swap in their own frontier model.
- SE3 plans to grow "from roughly 40 employees to about 120". Fortune gives no contract values for the Bundeswehr or Ukraine work and no measured performance figures; the maturity assessments are the CEO's own, and one outlet has the story.
Health, science & medicine
Google DeepMind releases AlphaGenome 2, which it says finds nearly twice as many expression-changing variants beneficialCompany claimPreprint
- DeepMind says AlphaGenome 2 "identifies nearly twice as many variants that change gene expression in cells, compared to our previous state-of-the-art model", and was "trained on more than 100 million individual cell profiles across 68 tissues". Unlike AlphaGenome 1, it is a conditional model that takes a tissue, cell type or cell state as input alongside "a 1-million-letter DNA sequence", and it adds DNA methylation to the readouts it predicts.
- Nature reports that DeepMind replicated aspects of a 2024 finding on the ETS2 gene in inflammatory bowel disease. James Lee of the Francis Crick Institute, who led the 2024 study, told Nature: "They've managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments."
- The model is "available today for non-commercial research via the AlphaGenome API, and in private preview for commercial research".
- The DeepMind post gives two different figures for the training set — "more than 100 million" and "over 107 million" individual cell profiles — and Nature says the model is described in a preprint on the company's website rather than in a peer-reviewed paper. The near-doubling claim is DeepMind's own and has not been independently evaluated.
Nature paper: a latent-diffusion imaging method independently reproduces the M87* black hole ring beneficial
- The paper, "Independent interferometric observable confirms black hole shadow in M87*", reports a reconstructed "ring-like morphology, with an angular diameter of 42.2 ± 1.5 μas corresponding to a black hole mass of 6.5 ± 0.5 billion solar masses assuming nominal scaling parameters, observed at an inclination angle of approximately 20° from the approaching jet".
- The method, "Generative Deep learning Image Reconstruction with Closure Terms (GenDIReCT)", is described as "unique in using closure invariants to condition a latent diffusion model", combining a conditional U-Net latent diffusion model with a multilayer CNN. A complementary Bayesian forward-modelling fit gave "a mean diameter and standard deviation of d = 42.3 ± 1.1 μas", which the authors call "compatible with" the image-reconstruction result.
- The striking methodological detail is that the training data is not astronomical: "models described in this work are trained on augmented non-astronomical images from the CIFAR-10 dataset, as well as simple geometric shapes", and the authors say training without the geometric shapes produced "no notable difference in the image reconstruction". That weakens the usual worry that such reconstructions encode a ring the training set taught them to expect.
- Compute and inputs are stated: "we run the imaging sequence for 96 total GPU hours on a single NVIDIA H100 GPU for each dataset, producing several hundred (about 300) individual image reconstructions, each drawing from 4,096 diffusion samples", using publicly available 1.3-mm Event Horizon Telescope measurements from four 2017 epochs, two 2018 epochs and two 2021 epochs.
Preprint: billed AI use in US care rose nearly tenfold since 2021 but reached only 2.9 per 100,000 encounters mixedPreprint
- The serial cross-sectional study covers Epic Cosmos data on 11,605,867,309 encounters among 233,628,292 US patients from January 1, 2021 to April 14, 2026, and finds "335,349 encounters (2.9 per 100,000) included an AI-required CPT code" out of 65 codes examined.
- Growth was steep from a very low base: "the odds of an encounter with an AI-required code were 9.88 (95% CI, 9.71-10.06) times higher in 2025 than in 2021". The authors describe billed AI use as having "remained rare but grew substantially".
- Unadjusted odds ratios point to uneven access: higher for Medicare (OR 5.11) and Medicare Advantage (OR 5.12) than commercial insurance, and lower for Black patients (OR 0.72), Hispanic or Latino patients (OR 0.34), Medicaid enrollees (OR 0.65), residents of the most socially vulnerable areas (OR 0.53) and the most rural areas (OR 0.86).
- The study measures billing codes that require AI, not AI use in general, so ambient documentation, triage tools and anything not separately billable are outside it. The odds ratios reported here are unadjusted, and the paper is a preprint by authors at UC San Diego, UCSF and Seattle Children's.
Google Research says coaches picked its world-model-grounded wearable health plans in 71.4% of blinded reviews beneficialCompany claimSingle source
- Google Research describes a "coachable health age" model and SensorWM, "an action-conditioned world model pre-trained on 2.28 million person-weeks of de-identified data from roughly 100,000 consenting Fitbit users". In a blinded head-to-head study with "five certified lifestyle coaches and five board-certified primary care and lifestyle medicine physicians across 49 participant profiles spanning five behavioral personas (490 paired reviews total)", it reports that "lifestyle coaches chose the counterfactual-grounded plan as the best overall plan in 71.4% of evaluations (p < 0.001)" and rated it more likely to drive long-term adherence in "71.0% of evaluations (p < 0.001)".
- On safety, Google says physicians "rated 99.2% of counterfactual plans within safe exercise intensity limits and 97.6% accurate on physiological claims".
- Validation cohorts cited include the All of Us research program at n = 22,748 with "more than 68,000 person-years" of Fitbit and electronic health record follow-up, a longitudinal Fitbit cohort at n = 1,464, and a multimodal cardiovascular-kidney-metabolic model evaluated in 819 WEAR-ME participants with "modified Life's Essential 8 r = −0.58; metabolic syndrome severity r = 0.65".
- Google says the models are "investigational research prototypes rather than clinical diagnostic tools" and that the evaluations were offline and retrospective. No patient outcomes were measured — the study scored plans as judged by clinicians, not health results from following them. The figures are Google's own, in one post.
Nature: AI search summaries are resurfacing a palaeontologist's discredited papers, none of them retracted harmfulSingle source
- Nature reports that papers by palaeontologist Vishwa Jit Gupta "are being surfaced through artificial-intelligence tools that summarize search results", and that for searches on Himalayan landmarks such as the Tandi Limestone "Google's AI overview obliges with direct links to Gupta's discredited papers".
- The scale of the underlying problem: Nature says "littered through more than 450 publications were maps of collection sites that didn't exist, plagiarized images and specimens stolen from collections", and "more than 100 experts had been duped into co-authoring descriptions of fossils". A 1991 Journal of the Geological Society of India evaluation of 19 of his papers concluded "the fossil finds of V. J. Gupta are not reliable [...] the data is incomplete bordering on disinformation".
- The mechanism is a gap in the record rather than a model failure: Nature says none of Gupta's papers have been retracted, so the literature carries no machine-readable signal that would let a summariser exclude them.
- Chris Graf, Springer Nature's director of research integrity, told Nature that where investigations of old papers are hampered by limited information, an editorial expression of concern "is likely to be our preferred mechanism for alerting readers to potential issues". Nature notes its news team is editorially independent of Springer Nature. Nature does not say how often the AI overviews surface these papers or across which queries; one outlet has the story.
Policy, regulation & law
European Commission proposes a Digital Fairness Act that would ban interfaces forcing consumers to use automated systems
- The Commission published COM(2026) 1100 final, a "Proposal for a Digital Fairness Act", on 11 November 2026. Tech Policy Press reports it would repeal the Unfair Commercial Practices Directive and replace it with a regulation covering business-to-consumer commercial practices online and offline, with Chapter III, "Digital fairness", running Articles 10 to 29.
- The provision that bites on AI agents is Article 26. Tech Policy Press reads Article 26(9) as an outright ban on interfaces that require consumers to use automated systems to access or conclude transactions, while Articles 26(1) to (4) would make automated contracting tools offer parameter controls, have the consumer choose a confirmation level on first use, and disclose payments and conflicts of interest. Article 26 does not apply to financial services.
- Article 12 would ban targeted advertising to minors where the trader is aware "with reasonable certainty" the consumer is a minor, and ban profiling using special categories of personal data under GDPR Article 9(1), with no harm threshold; Recital 59 extends personalised recommendation rules to "natural language systems". Article 13 would conditionally ban personalised pricing based on profiling that targets minors or vulnerable consumers where harm is reasonably likely, with Recital 60 exempting personalisation that benefits consumers, such as student rebates. Article 10 prohibits dark patterns, with the banned practices listed in Annex I.
- Article 30 would give the Commission direct enforcement powers, but Tech Policy Press says the proposal sets no EU-level fine amounts, leaving penalties to member states, which the authors call undesirable. This is a proposal at the start of the legislative process, not law; the article-level readings above are Tech Policy Press's analysis of the text rather than Commission summaries.
EDPB picks AI-driven automated decision-making in recruitment for its sixth coordinated enforcement action
- At its plenary in Brussels on 11 November 2026, the European Data Protection Board "selected the topic for its sixth coordinated enforcement action (CEF)", which "will concern automated decision-making (ADM), including AI-driven ADM, in recruitment".
- The EDPB says the action "will focus on the enforcement of the right not to be subject to decisions based solely on automated processing and the related right of access, in the context of the use of automated candidate screening tools, including AI-driven ones" — so it targets GDPR Article 22 rather than the AI Act.
- Timing: "DPAs will join this action on a voluntary basis in the coming weeks, and the action itself will be launched in the course of 2027." IAPP notes the EDPB has issued no ADM-specific work since the Article 29 Working Party's May 2018 guidance on Article 22, and that the final report on the separate 2026 transparency action is not expected until early 2027.
- The same plenary adopted a statement on the first review of the EU–Republic of Korea adequacy decision, noting that Korea's PIPA was amended to introduce "a right to refuse an automated decision (incl. by AI systems) or request an explanation if the decision significantly impacts the data subjects' rights or obligations". Nothing is enforced yet: participation is voluntary and the action begins next year.
Analysis of 56 amicus briefs in the OpenAI copyright MDL counts 798 case citations before 20 November replies Single sourceUpdate
- Tech Policy Press published a dataset and analysis of the amicus filings in In re: OpenAI, Inc. Copyright Infringement Litigation before US District Judge Sidney H. Stein in Manhattan (CourtListener docket ID 69879510). Judge Stein accepted 56 amicus briefs on Nov. 6, 2026, filed between Oct. 14 and roughly Oct. 30 with most on Oct. 29 or 30, under a 6,000-word limit.
- The declared positions split 27 supporting plaintiffs, 18 supporting OpenAI and Microsoft, 7 declaring no side and 4 neither supporting nor opposing. The analysis counts 798 case citations across the briefs, with the most-cited decisions being Campbell v. Acuff-Rose (1994, cited by 44 briefs), Authors Guild v. Google (2015, 39 briefs) and Warhol v. Goldsmith (2023, 36 briefs).
- Procedurally, replies are due Nov. 20, 2026 and no oral argument date is set. The Justice Department filed a statement of interest on Sept. 1 backing OpenAI, and Reps. Raskin, Nadler and Johnson filed in response arguing it deserves little weight.
- Tech Policy Press says the citation analysis "was conducted with Anthropic's Claude Opus 5.5 under editor-set rules", with ambiguous citation decisions not reviewed by an editor but logged in downloadable data — so the 798 count is machine-generated and partly unaudited. The briefs and the Nov. 6 order predate this window; the dataset and analysis are what is new.
EU leaders hold their first leaders-level AI discussion in Dublin on 13 November, with Mistral's CEO and Hinton presenting Single source
- Tech Policy Press reports that an informal meeting of EU Heads of State or Government in Dublin on Friday, November 13, hosted by the Irish Government, has an agenda "focused on AI" and is to be attended by 27 EU member state leaders. Council President António Costa's invitation letter describes it as the first discussion "at Leaders' level" on AI at the Council.
- The agenda has two parts: European competitiveness and AI sovereignty; and AI risks, safety and security plus European contributions to international governance of advanced AI. It includes presentations by the CEO of Mistral AI and by Geoffrey Hinton.
- The meeting follows a European Political Community summit in Dublin on Thursday, November 12, themed "Europe – strong, secure and connected", with more than 50 world leaders expected and about 5,000 Gardai — roughly a third of the national force — working the event.
- This is an agenda, not an outcome: informal European Council meetings produce no binding conclusions, and nothing has been decided. One outlet is the source for the agenda and the presenter list.
Compute, chips & infrastructure
Nvidia's board met Wednesday on Jensen Huang's future after reports it is preparing to remove him; shares fell about 9% Update
- Two sources told NBC News the board was addressing Huang's future and that the meeting took place Wednesday evening, with one saying "the White House was aware". Nvidia shares "dropped almost 9% in after-hours trading"; NBC says that if the drop held through Thursday's opening bell it "would erase more than $350 billion in value". TechCrunch puts the after-hours fall at "about 8%" and notes the stock is down about 11% over the past five days.
- Fortune, citing the Wall Street Journal, says the board is "preparing to remove Chief Executive Jensen Huang" and "could meet as soon as Wednesday to determine his fate"; it reports after-hours shares "down more than 7%" at the time of writing. CBS News, citing two sources, says the board "is preparing to oust CEO Jensen Huang soon". Bloomberg reported, per Fortune and TechCrunch, that Nvidia is working with an outside firm to identify CEO candidates; Fortune names CFO Colette Kress and lead independent director Stephen Neal as possible interim CEOs, attributing those names to experts it quoted rather than to the Journal.
- The trigger inside this window was a second report. The New York Times, citing five people with Nvidia experience, reported that Huang had a relationship with an Nvidia executive who reports to him, on top of the affair with Alice Huang — no relation — whom US intelligence reportedly assessed to be a Chinese spy and who previously led Megaspeed International, a Singapore firm that bought billions of dollars of Nvidia chips. NBC says it has not confirmed the Times report.
- No decision has been announced. Nvidia's board said on Sunday it had "confidence in Jensen and his leadership team" and "no reason to believe that either national security or NVIDIA secrets were compromised". The ouster reporting rests on anonymous sources at the Journal, Reuters, CBS and NBC; this briefing could not open the Journal or Times articles directly and is relying on these four outlets' accounts. The briefing covered the board's investigation on 11 November.
CoreWeave expects about $9 billion in customer prepayments and reports $14.5bn of Q3 capex Update
- TechCrunch reports CoreWeave expects about $9 billion in prepayments and deposits over the next three months, more than $3 billion of it tied to signing new contracts for capacity it has already built and so requiring no additional hardware spending. CEO Michael Intrator said the company "will not use the unsecured debt market to fund our business at current yields"; some of its bonds yielded above 15% last month.
- The pricing figures show why short contracts are attractive: prices rose about 25% in July and another 10% later in the quarter, with short-term capacity selling for more than $40 million per megawatt annually against $20 million to $25 million per megawatt annually for new infrastructure on four- and five-year contracts.
- DCD reports Q3 capex of $14.5bn, adjusted operating income of $379 million against prior guidance of $200m to $260m, and a net loss of $709 million versus a $110m loss in the same quarter of 2025; TechCrunch gives quarterly interest expense of $846 million. Intrator said the average cost of debt across the business is 7.9 percent and that "through 2027, more than 75 percent of committed capex is tied to investment-grade and enterprise customers". CFO Nitin Agrawal said more than 50 percent of the $132.5bn backlog "is attached to a contract where we expect customer delivery to commence by the end of this year".
- The $9 billion prepayment expectation is guidance, not booked cash, and the shift to shorter contracts trades duration risk for price. The briefing covered Q3 revenue of $3,661 million and the $132.5bn backlog on 11 November; the funding, capex and per-megawatt pricing figures are what is new.
UK chip startup Fractile announces a $720M Series C; the British government cites a valuation over £5 billion Single sourceUpdate
- TechCrunch reports Fractile raised a $720 million Series C led by Thrive Capital, Founders Fund, Lightspeed Venture Partners and Redpoint, with Britain's Sovereign AI Fund also backing the round. The British government said the valuation is "over £5 billion"; TechCrunch says this is "quite possibly about $7 billion", a figure reported by the Financial Times and a Reuters source.
- The UK government said the goal is "ensuring the U.K. has a greater stake in the hardware that will underpin its future economy" — state money going directly into an inference-chip challenger rather than into compute capacity bought from US vendors.
- On the technology, TechCrunch says Fractile's design stacks DRAM on top of each chip and the company takes a full-stack approach including an ML compiler. CEO and founder Walter Goodwin said the funding will let the company "ship [its] first chips and systems next year".
- TechCrunch gives no benchmark or performance numbers for Fractile's silicon, and no chips have shipped. Sifted and the Telegraph reported a smaller round of about £540m and Sovereign AI Fund participation in late October and on 1 November; the $720M figure and the government's £5bn valuation statement are what appeared on 11 November. One outlet carries the detail.
SpaceX seeks a Starbase tax abatement for a gas plant and supercomputing site, investing up to $24 billion Single source
- DCD, citing a new public notice reported by Bloomberg and others, says the Starbase City Commission is scheduled to meet on December 9 to consider approval of a property tax abatement agreement with SpaceX. The gas-fired power plant and supercomputing facility would sit on a 444-acre tract just east of Brownsville, between Boca Chica Blvd and the Port of Brownsville, in Cameron County, and SpaceX "could invest up to $24 billion in the project", built out in phases.
- The pattern is now familiar: an AI compute site paired with its own gas generation, sited to avoid grid interconnection queues, and brought forward with a local tax abatement.
- DCD notes SpaceXAI already runs a cluster of datacenters around Memphis — one in Tennessee, the rest in Mississippi — used to run Grok and to sell AI cloud services to clients including Anthropic and Google, and that Musk recently announced two new Mississippi datacenters named Gun Range and Spam and claims more than 1GW of capacity in the Memphis cluster, "though these claims have been disputed". The company is also looking at a 6GW cluster in Cochise County, Arizona.
- Nothing is approved: the commission vote is scheduled for December 9 and the full scope has not been shared. The $24 billion is an upper bound in a public notice, not committed spend, and one outlet carries the story.
Deployment & impact
United Airlines bans third-party personal AI agents from booking flights, citing credential sharing and federal rules mixedSingle source
- CBS News reports that United said on Wednesday it is banning AI agents from making flight reservations for travelers, and is updating its booking and ticketing policies to match. United said "sharing account credentials with personal, third-party, AI assistants can expose sensitive information" and that "actions taken by these tools may result in unintended purchases, account changes, or use of MileagePlus miles". It said it is "taking steps to restrict access" to "prohibit use by unauthorized, third-party personal AI agents or assistants", both on its own digital platforms and indirectly through travel agencies and aggregators.
- United also said that using an AI agent to check in for a flight could violate federal rules requiring travelers to verify in person that their luggage contains no banned substances — a compliance argument rather than a commercial one.
- This is a large consumer platform shutting out agents by policy rather than fixing the authorisation model that makes them risky. CBS names Meta's Muse and OpenAI's dots among the assistants at issue, contrasts United with Southwest, which has made its flight inventory viewable to ChatGPT users, and says United fares will remain bookable through Kayak, Hopper and Google Flights. Henry Harteveldt of Atmosphere Research Group told CBS the policy suggests United wants to keep control of the booking process and avoid the cost of AI bot flight searches.
- CBS gives no effective date, no enforcement mechanism and no figures on how much agent traffic United sees. The article names no United spokesperson, attributing the policy statement to "the airline", and one outlet carries it.