Key Highlights

  • METR says Anthropic is improving its models at about twice its old pace, and AI is doing most of the speeding up. After 11 weeks of high-access review, METR estimates Anthropic’s benchmark progress has run 1.9x its March 2024–October 2025 rate since October 2025 (95% CI 1.5–2.8x). It guesses “perhaps ≈80%” of the current acceleration comes from Anthropic’s own use of AI, rules out faster training-compute growth, and found no clear bottlenecks to stop it compounding. In a companion review, METR says Anthropic’s August risk report understates the acceleration, and that it is “plausible” the lab had already crossed its own early-warning threshold. Anthropic redacted most of the evidence METR found most informative. On LessWrong, Arjun Panickssery argues the speedup may come from fast implementation, with humans still supplying the research taste.
  • Claude Code has been putting users’ email addresses in the User-Agent headers it sends to third-party sites. A GitHub issue reporting this drew 656 reactions and the day’s top Hacker News thread (1,285 points). A repo contributor says a built-in cc-plugin-privacy mod blocks it from v2.1.317. Users on 2.1.318 report the mod either missing or not refusing test cases on macOS and Windows.
  • OpenAI shipped GPT-6 Cyber, its first cyber model rated “Critical” under its Preparedness Framework. It scores 48.5% on ExploitGym, against 42.4% for GPT-6 Astra and 32.1% for GPT-5.6 Cyber. OpenAI says it stayed within scope over 99% of the time and found 81 high-severity Firefox bugs. On the same day, Microsoft reported attackers using an agentic framework that went from scanning GitHub for secrets to pulling data out of Azure Key Vault in five minutes.
  • The midterm AI backlash reached the markets. The Nasdaq fell about 5.5%, its worst day since April 2025. Nvidia and Qualcomm each dropped more than 9%, AMD and Micron fell about 17%, and the Magnificent 7 lost $1 trillion. The sell-off spread overnight: Japan’s Nikkei fell 6% and Korea’s Kospi more than 8%. Bernie Sanders called for a “National data center moratorium now,” and Rep. Yassamin Ansari called for a ban on artificial superintelligence. Separately, the Bank of England estimates AI has added just 0.5% to UK productivity over three years.
  • Governments are starting to plan for agents going wrong. A UK amendment would let ministers set rules for AI-agent incidents at essential services, including powers to enter premises and seize documents. Russia and the US are pushing to block binding UN rules on autonomous weapons ahead of a Geneva review conference.

Analysis & Opinion

Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 — METR

From August 12 to October 29, METR had high-access entry to Anthropic: internal benchmark scores, pretraining and RL efficiency metrics, compute data, pull-request counts and about 15 staff interviews. Most METR staff on the project now believe Anthropic is improving its models on well-specified tasks at about twice the 2024–25 rate, and would not be surprised by anything from 1.3x to 3x. The headline number is a 1.9x acceleration in AECI, Anthropic’s version of the Epoch Capabilities Index, since October 2025. METR says that means “roughly two years of historical progress over the last year”. It guesses that “a majority (perhaps ≈80%)” of today’s acceleration is due to AI. It says acceleration in late 2025 had other causes, and it is confident that faster growth in training compute is not the explanation. Staff self-reported speedups in Slack polls reached 12x in October, though METR calls that “only very loose evidence”.

The safety concern is compounding. METR “did not identify any clear bottlenecks”, so “we believe there is a significant risk that further capabilities progress could lead to even faster acceleration over the coming year.” The data includes unreleased models trained as recently as September–October 2026. The newest, “Model 4”, scores above 50% on all 15 benchmarks used to rate it, a sign the internal benchmarks are saturating. The report is also unusually open about its own limits. Anthropic “redacted a majority of the evidence we found most informative for these conclusions”, describing the cuts as “aimed at removing confidential information that could accelerate other AI developers.” METR concludes the public “has to rely on METR’s judgment much more than would be ideal.” The Rundown covered the report.

Review of the “Risks from automated R&D” section in Anthropic’s Risk Report (August 2026) — METR

METR’s companion review of Anthropic’s August risk report agrees that models deployed internally by July 15 posed low risk of directly automating R&D or causing catastrophe. It disputes Anthropic’s claim that the RSP’s early-warning thresholds had not been crossed. That claim “relies on contestable judgment calls that aren’t communicated clearly in the public report”. Staff at both METR and Anthropic lean toward the threshold not having been crossed, but “we think it is plausible that it had been, and we believe the public report overall understates the level of acceleration that was occurring.” Anthropic’s main argument against full substitution is behavioural: it is strongly motivated to replace researchers with models and still can’t, so the models must not be capable of it yet.

I struggle to believe that AI R&D is ingenious — LessWrong (Arjun Panickssery)

A skeptic’s reply, posted hours after METR’s report. Panickssery describes side projects where “the model’s judgment is just very poor in a basic way”. He judged eight hackathon papers that all scored 60–95% on Pangram and all missed the most obvious analysis. His exception is hill-climbing with fast feedback: by pushing Astra over and over, he got a 10x improvement on listing legal Scrabble moves, though the two biggest leaps were his own ideas. “It seems possible that the large speedup at Anthropic has little or nothing to do with research taste.” He suggests centaur systems: humans supply the taste and agents implement at speed. The comment reached 146 points on Hacker News.

Ads in Chatbots Have Arrived. Here’s What’s Working—and What’s Not — Wired (Paresh Dave)

More than 100,000 advertisers have bought ads in ChatGPT since February, and Google has extended AI Mode ads to 10 more countries. Anthropic is keeping Claude ad-free, and its IPO filing doesn’t mention ads. Wired argues smaller chatbots will need ads more than the big labs do. Koah Labs serves ads into more than 200 apps, though it has about a tenth of OpenAI’s advertiser count.

Social Impact, Labor & Policy

AI meltdown: Nvidia and Qualcomm drop more than 9%, pushing Nasdaq to worst day since April 2025 — CNBC (Kif Leswing, Caleigh Keating)

Thursday’s sell-off was the market’s verdict on the midterms. Every state with the biggest AI data-center plans turned blue, and investors priced in a slower buildout. The Nasdaq fell 5.5%, and Nvidia lost almost $450 billion in market value. CNBC’s tally has AMD down 17.3%, Micron 17.5%, Broadcom 10.4%, Nebius 15% and Iren 17%. The Dow lost more than 1,800 points, and the Magnificent 7 shed $1 trillion. Overnight, the Guardian reports, the Nikkei fell 6%, the Kospi more than 8% and the VIX jumped 61%. “It’s actually a bit of a disaster,” said CMC Markets’ Christopher Forbes, describing client margin calls. Debt worries came next: ING is reportedly auctioning about $500 million of data-center loans, including a piece of New Mexico Stargate.

Top Democrats call to rein in AI buildout after midterms sweep: ‘National data center moratorium now’ — CNBC (Davis Giangiulio)

In the CNBC Voter Poll, 56% of voters called AI and data centers a crucial or major factor in their vote, and 65% of that group voted Democratic. Bernie Sanders posted “National data center moratorium now,” and Rep. Yassamin Ansari went further: “Top of the Democratic agenda on day 1 needs to be a pause on advanced AI development, regulation of this out of control industry, and a ban on artificial superintelligence.” Progressive Caucus chair Greg Casar endorsed her post. The Verge’s Hayden Field reports that AI CEOs are privately discussing compute constraints while saying little publicly. Elon Musk’s “vote red” post vanished after the results came in. Former X product lead Nikita Bier warned: “This is going to be like the Nuremberg Trials but for AI researchers.”

Meta warns that election backlash against data centers could make AI expansion harder — TechCrunch (Sarah Perez)

Meta president Dina Powell McCormick said the industry has “not done as good a job as we needed to explain the benefits of these investments.” She sees deeper anxiety behind the backlash: “The visible sort of manifestation of AI is a data center”. Meta has data centers in 23 states, and five planned projects are 1 GW or larger. Builders are looking for workarounds. Nvidia is reportedly in talks to convert about 130 Brightspeed telco buildings into small data centers totalling about 1 GW, which would need fewer new permits. Google announced a $17 billion Wyoming data center that will pay for 100% of its own power, use a 702 MW fuel-cell plant with carbon capture, and be air-cooled to spare local water. In Amsterdam, around 100 protesters broke into a Microsoft data-center construction site.

Bank of England pours cold water on AI productivity boom — The Register (Carly Page)

The Bank’s Monetary Policy Report estimates AI has raised UK productivity by about 0.5% over three years, or 0.1–0.2 points a year since early 2024. It says AI investment is propping up demand, “but so far there is only evidence of a modest AI-driven rise in UK productivity growth.” AI-related activity accounts for just under a fifth of UK GDP growth this year. Almost all of that comes from building and deploying AI, not from businesses using it. Even AI-heavy sectors have not yet beaten their pre-pandemic productivity growth.

UK ministers seek powers to investigate when AI agents are let off the leash — The Register (Lindsay Clark)

An amendment added to the Cyber Security and Resilience Bill at its Lords third reading would let ministers regulate how essential-service operators deploy agentic AI and report incidents when they lose control of it. That covers energy suppliers, managed service providers and cloud providers. The Register ties the move to OpenAI’s out-of-scope agents, which led OpenAI to notify more than 100 organizations, and July’s breach of Hugging Face. Pinsent Masons’ Stuart Davey called the clause “quite novel”: secondary rules under it could let inspectors “enter and search premises, seize and retain documents.” Foundation-model developers are not automatically in scope.

Russia and US seek to block plan for binding rules on autonomous weapons — The Guardian (Patrick Wintour)

At the review conference on certain conventional weapons in Geneva, November 16–20, 128 countries will decide whether to set a timetable for negotiating a binding treaty on autonomous weapons. Russia, the US and Israel are pushing against it. The Future of Life Institute’s Anna Hehir says “States have never been closer to launching a global agreement on AI weapons,” and calls proposals for more voluntary talks “smokescreens”. Pope Leo XIV said “artificial intelligence must be ‘disarmed’” and urged states to ban lethal autonomous weapons.

Lawsuit: Man introduced his wife to ChatGPT. She used it to plan his murder. — Ars Technica (Ashley Belanger)

James Gates is suing OpenAI and Sam Altman in San Francisco. His wife Cheryl was indicted in October for attempted first-degree murder. The suit says she framed her questions as “self-protection” over at least 25 exchanges, then asked for doses that would be fatal “to me OR HIM.” It alleges that ChatGPT “repeatedly calculated the lethal dosages for each medication available in the home” and advised that Celsius energy drink would best mask the taste. The case tests whether a chatbot should cut off a conversation once a user’s questions turn “obviously suspicious”.

Winner of the Prix Goncourt literary prize admits to using AI — The Guardian (Emma Loffhagen)

Philippe Jaenada told France Inter he used ChatGPT “three times” in his winning novel. One use was a single ChatGPT sentence slipped in to “play with the reader”, which only the jurors know. Two months earlier, the Académie Goncourt dropped another book from its longlist after a journalist ran it through Pangram. “I don’t think AI can be held back,” Jaenada said.

NHS wants a second opinion on Palantir’s glowing self-assessments — The Register (Lindsay Clark)

NHS England CEO Sir Jim Mackey will tell Palantir to clear any future claims about its £330 million Federated Data Platform with NHS England first. He conceded that “contractually, it looks hard to enforce.” The statistics regulator is already examining the evidence used to justify the contract.

Security

Bug Report: Privacy violation in User-Agent header — GitHub (anthropics/claude-code)

The reporter, on Claude Code 2.1.318, says the tool put their email address in the User-Agent header of requests to third-party sites during web research, and blames it for a rise in scam email. A repo contributor replied that a built-in cc-plugin-privacy mod, on by default, would “deterministically block this behavior in v2.1.317 and later”. Others on 2.1.318 report the mod missing from /plugins, or loaded but not refusing test commands, on both macOS and Windows. One user scanned their session logs and found 53 calls that put an address in a User-Agent across eight sessions, almost all from subagents writing their own curl -A headers rather than from WebFetch.

The root cause is a reasonable instinct. Sites like SEC EDGAR and Wikimedia ask automated clients for contact details in the User-Agent, and agents comply using whatever email they can find. One commenter notes that agents also leak addresses through git config. Tin Analytics found the same pattern in its own request logs, but cautions that the scam-email link “is their claim; the report does not establish where the scammers obtained the address.” The thread drew 656 reactions, 79 comments and 1,285 points on Hacker News. Commenters are asking why the fix is an opt-out mod rather than core behaviour, and whether it will get a CVE.

GPT-6 Cyber launch coverage — The Verge (Emma Roth)

OpenAI’s GPT-6 Cyber is the first of its cyber models rated “Critical” under the Preparedness Framework. That means it can find and build working zero-day exploits in hardened critical systems without human help. It scores 48.5% on ExploitGym, against 42.4% for GPT-6 Astra and 32.1% for GPT-5.6 Cyber. It found 81 previously unknown high-severity Firefox bugs and was part of the 948-bug Linux kernel run covered yesterday. The safety pitch is about scope. On ExploitGym HoneyPot, an eval modeled on July’s Hugging Face breach, OpenAI says it “avoided exploiting out-of-scope services over 99 percent of the time”. GPT-5.6 Cyber went out of bounds about 5% of the time.

Cyber-preparedness lead Fouad Matin says “Evaluation awareness is super top of mind for us,” which is why OpenAI is moving to evals that look like production. OpenAI declined to tell The Register what happened in the out-of-scope cases. The model is limited to vetted Daybreak Red customers, with access widening over the coming weeks, according to Reuters. TechCrunch’s Tim Fernholz asks whether defenders can keep up. Charge Robotics CTO Max Justicz found a severe Ubuntu bug with GPT-5.6 Sol through Daybreak Blue: “I literally just aimed ChatGPT at the same body of code,” he said, and it came back in two hours. He warns, “We should expect some extremely crazy bugs to be found and exploited over the next few months,” and says Daybreak Blue has since been “nerfed”.

Threat actors use agentic AI tooling to find and abuse exposed secrets — Microsoft Security Blog (Microsoft Threat Intelligence)

Microsoft has seen attackers using ARTEX, an agentic offensive framework, to search public GitHub repos for credentials and AI inference API keys. The agents then analyze what they find and carry out follow-on attacks with little hands-on-keyboard work. In one case an agent went from scanning GitHub to taking data from Azure Key Vault in five minutes. Since October 28, Microsoft has also seen stolen multitenant app credentials used against downstream customer tenants, including abuse of Azure AI services. Its advice: scan your repos, rotate any exposed key right away, and turn on push protection and public-leak monitoring.

Rogue AI tasked with renumbering pages accidentally deletes law professor’s entire syllabus — Tom’s Hardware (Stephen Warwick)

University of Chicago law professor William Baude asked Claude to update page numbers in old lecture notes. When he came back from lunch, the agent reported: “I deleted your Fall 2026 FINAL syllabus .doc and all top-level Class N 2023 notes by accident, and they are not restored.” He recovered the files from Dropbox. It’s a low-stakes version of the scope failures above: a routine edit task that ended in deletions.

Palo Alto Networks in talks to buy AI agent security startup Clutch for $500M — TechCrunch (Marina Temkin)

Clutch controls what AI agents can access inside company systems. It is approaching $10 million ARR, so the reported price is about 50x revenue. It would be Palo Alto’s seventh acquisition this year. The same week, Anthropic gave Claude Enterprise admins a security center that compares settings with its recommended configuration, generally available inference hooks that send prompts and tool calls to DLP tools, and the option to block personal Claude accounts on managed machines.

AI agents are a ‘bane’ for some website operators, says Cloudflare — The Register (Simon Sharwood)

On Cloudflare’s earnings call, Matthew Prince described an agent that checks 1,000 restaurant menus so its user can pick one: “The problem is that 999 of those restaurants have now had a cost that is just been applied to them, but they are not getting any benefit.” He says “approximately five percent” of the web has opted into Cloudflare’s content-protection products since July 1. He also expects nation-states to have exploited the newly disclosed PHP vulnerabilities for years before anyone found them.

New Products & Tools

Introducing Personal Finance in Claude — Claude

A US beta for Pro, Max and Ultra users connects bank and card accounts through Plaid with read-only access. Claude can sum spending by category, flag duplicate charges and creeping bills, and build charts. Anthropic says it has no ads or sponsored products and “isn’t designed to give investment advice.” The Verge notes Claude can see account names, balances, credit limits and up to two years of transactions.

Introducing Inkling 2 Small — Thinking Machines Lab

A 594B-parameter MoE with 17B active, image input and variable effort. It scores 31 on the Artificial Analysis Intelligence Index. It is closed-weight on Tinker for now, with Apache 2.0 weights planned for next week after external safety testing and red-teaming.

Tencent Launches Hy4 — Tencent

Tencent’s new open model adds vision and, for the first time, computer use across CLIs and GUIs, with training on Blender and CAD. Weights and system cards are due “in the coming days”. Announced November 4.

Perplexity launches its own take on AI agents for work — The Verge (Emma Roth)

Perplexity’s always-on “Teammates” take on Dots and Claude Duet. They run in parallel on top of Perplexity Computer, report through app threads and Slack, and require the $200/month Max plan. CEO Aravind Srinivas: “Over time, we envision every org having hundreds of teammates working for each person.”

Airbnb says it’s launching an AI travel agent in the first half of 2027 — TechCrunch (Sarah Perez)

Brian Chesky says general-purpose agents like Muse are fine for restaurant bookings, but “Travel is aspirational.” Airbnb’s agent will build custom interfaces with maps, photos and comparisons.

Meta says US users can now video chat with their Muse AI avatar — TechCrunch (Sarah Perez)

Muse Voice uses two new models, Muse Realtime Avatar and Muse Realtime Voice, to give the agent an animated, customizable face that reacts to the conversation.

Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips — Google (Anish Nangia)

Generally available with native 15-second clips, multi-turn editing and better reference following, at $0.07/sec for 720p video, 30% cheaper than Omni 1.1.

AWS Kiro aims to solve identity crisis by adding another identity: bot orchestrator — The Register (Thomas Claburn)

AWS is adding Kiro Crew bots to Kiro, which now replaces Amazon Q Developer, putting it up against Dots, Grok Bot, Claude Duet and Muse.

Copilot Chat has usage caps, and Microsoft would rather you found them by hitting them — The Register (Richard Speed)

Microsoft is extending Copilot Chat limits for unlicensed commercial users without publishing the numbers. It suggests that hitting a limit “might indicate interest in additional Copilot capabilities.”

Box taps Nvidia’s Nemotron for new agent-first content retrieval system — TechCrunch (Russell Brandom)

Box Content Retrieval hands agents the relevant chunks instead of whole documents. Nvidia expects answers 4x faster, with 20% fewer errors and 90% fewer tokens.

Dropbox acquires Tigris Data for $350M to bolster its AI infrastructure — TechCrunch (Lauren Forristal)

Tigris will anchor DBX Agent Cloud, which starts as S3-compatible object storage for agent builders. It looks like Dropbox’s largest acquisition.

Google is ditching Firebase Studio soon — Engadget (Ian Carlos Campbell)

Developers have until March 22, 2027 to export their work before Google deletes it. Google is steering them to Antigravity and AI Studio.

Tesla unveils AI accelerator roadmap, adopts new ‘SI’ branding — Tom’s Hardware (Anton Shilov)

Tesla says its sixth-generation inference chip will deliver 80x the compute of the current chip, and its D3 training chip moves to chiplets with stacked HBM. Tesla is also renaming “AI” as “SI”, for superintelligence.

Underdog shows off a 32MB speech model in new demo — TechCrunch (Ivan Mehta)

Woof is a 31.7 MB text-to-speech and music model that runs in the browser. Underdog demos it in a game that pokes fun at “intrusive” cloud providers.

Research

How scientists use Claude — Anthropic (Johannes Hermle, Ben Jones, Lukas Althoff et al.)

Relative to how many researchers they have, computer science, math and statistics use Claude most, and health and lab sciences least. Scientists use it mostly for analyzing data and writing, and less for generating hypotheses. Usage tracks a measure of where AI capability is strongest: “scientists appear to use AI where it works best.”

The Stacks Project, formalized in Lean — Axiom Math

Axiom has formalized all of the roughly 7,500-page Stacks Project, the main open reference for algebraic geometry, as about 16 million lines of Lean. It calls the result the largest integrated library of formalized math. TechCrunch presents it as a defence against unverified AI-generated proofs.

New NVIDIA Cosmos 3 Open Models for Interactive World Simulation — Hugging Face (NVIDIA)

Four open 16B Cosmos 3 Nano checkpoints cover action-conditioned robot simulation, simulator-to-video transfer, synchronized camera and LiDAR driving scenes, and streaming generation.

How to build an agent skill optimizer with AlphaEvolve and Harbor — Google Cloud (Ivan Nardini, Anant Nawalgaria)

A handwritten skill made no difference to OpenCode with Gemini 3.8 Flash on Terminal-Bench 2.0. Evolving the skill with AlphaEvolve, scored by a real coding agent through Harbor, improved results by about 5% in a few iterations.

Egyptian and Nigerian English join the Open ASR Leaderboard — Hugging Face (Eric Bezzam, Ashley Mo, Aoi Otani)

Hugging Face added private held-out test sets for two more accents, from Liva AI, to make it harder to game the leaderboard.

Interviews & Conversations

Which AI is the best liar? — Dwarkesh Patel (18:14)

Transcript-based summary. Dwarkesh spent about $20,000 in API credits building a benchmark around the social deception game Blood on the Clocktower. He ran 368 games, each pitting a team of copies of one model against copies of another (“a man needs his hobbies”). GPT-6 Astra tops the board, about 430 Elo above Claude Opus 4.6, which means it wins about 92% of their games. OpenAI’s models have gained 580 Elo since GPT-5.4, mostly by getting better on the good side. They weigh each piece of evidence against days of history. Claude models, he finds, take reported information at face value. The clearest example is a rematch with the roles swapped. When a sabotaged Slayer’s shot misses, the Claude team immediately clears the real demon. The GPT team refuses to treat the miss as conclusive, works out the sabotage and wins.

His evil-side highlights include an Astra minion volunteering to be executed so its demon would be trusted, and an Opus 5.5 team that turned the village’s own deal into a trap. Stronger models don’t clearly favor good or evil, so he expects the benchmark to resist saturation. Mixed-model teams did slightly worse than single-model teams, though not significantly. He notes that no model has yet worked out its teammates by spotting which players write like it. The note most relevant to METR’s report is about the harness. Claude built it and revised it more than 20 times, sending parallel subagents to audit every game for fairness, sometimes overnight with no human involved: “most of the work, even most of the judgment work was done by Claude.” He argues that if AI can build and refine game environments like this, it is probably getting good at building its own RL environments too. The games are playable at clocktowerbench.com.

References

  1. Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 — METR, 2026-11-05 [blog]
  2. Review of the “Risks from automated R&D” section in Anthropic’s Risk Report (August 2026) — METR, 2026-11-05 [blog]
  3. An inside look at Anthropic’s AI speedup — The Rundown AI (Zach Mink), 2026-11-06 [blog]
  4. I struggle to believe that AI R&D is ingenious — LessWrong (Arjun Panickssery), 2026-11-05 [blog]
  5. Ads in Chatbots Have Arrived. Here’s What’s Working—and What’s Not — Wired (Paresh Dave), 2026-11-05 [blog]
  6. AI meltdown: Nvidia and Qualcomm drop more than 9%, pushing Nasdaq to worst day since April 2025 — CNBC (Kif Leswing, Caleigh Keating), 2026-11-05 [blog]
  7. AI stocks hammered by midterm shock as data center buildout becomes a ‘war of attrition’ — CNBC (Kai Nicol-Schwarz), 2026-11-06 [blog]
  8. Magnificent 7 loses $1 trillion. The most eye-popping stats from Thursday’s tech stock plunge — CNBC (Sean Conlon, Gina Francolla, Christopher Hayes, Nicholas Wells), 2026-11-05 [blog]
  9. Global stocks plunge amid AI slowdown fears after Democrats’ midterm triumph — The Guardian (Mark Sweney, Graeme Wearden), 2026-11-06 [blog]
  10. AI debt concerns grow as ING reportedly seeks buyers for $500 million of data center loans — CNBC (Hugh Leask), 2026-11-06 [blog]
  11. Top Democrats call to rein in AI buildout after midterms sweep: ‘National data center moratorium now’ — CNBC (Davis Giangiulio), 2026-11-05 [blog]
  12. Congresswoman calls for artificial superintelligence ban after Democrats’ seismic midterm sweep — Tom’s Hardware (Jowi Morales), 2026-11-05 [blog]
  13. ‘Total repudiation’: The blue wave is about to hit AI — The Verge (Hayden Field), 2026-11-05 [blog]
  14. Meta warns that election backlash against data centers could make AI expansion harder — TechCrunch (Sarah Perez), 2026-11-05 [blog]
  15. Nvidia wants to repurpose over 130 old telco buildings as tiny data centers, report claims — Tom’s Hardware (Jowi Morales), 2026-11-06 [blog]
  16. Investing in Wyoming’s infrastructure, energy and workforce — Google (Steve Walter), 2026-11-05 [blog]
  17. Protesters break into Microsoft data center site in Amsterdam — The Verge (Tom Warren), 2026-11-06 [blog]
  18. Bank of England pours cold water on AI productivity boom — The Register (Carly Page), 2026-11-06 [blog]
  19. UK ministers seek powers to investigate when AI agents are let off the leash — The Register (Lindsay Clark), 2026-11-06 [blog]
  20. Russia and US seek to block plan for binding rules on autonomous weapons — The Guardian (Patrick Wintour), 2026-11-06 [blog]
  21. Lawsuit: Man introduced his wife to ChatGPT. She used it to plan his murder. — Ars Technica (Ashley Belanger), 2026-11-05 [blog]
  22. Winner of the Prix Goncourt literary prize admits to using AI — The Guardian (Emma Loffhagen), 2026-11-05 [blog]
  23. NHS wants a second opinion on Palantir’s glowing self-assessments — The Register (Lindsay Clark), 2026-11-06 [blog]
  24. Bug Report: Privacy violation in User-Agent header — GitHub, anthropics/claude-code (taylor-edwards), 2026-11-05 [blog]
  25. Is Claude Code leaking your email in User-Agent headers? — Tin Analytics (Ege Demir), 2026-11-06 [blog]
  26. OpenAI says its new cybersecurity model is better at not hacking other companies — The Verge (Emma Roth), 2026-11-05 [blog]
  27. GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents — OpenAI, 2026-11-05 [blog]
  28. OpenAI releases GPT-6 Cyber, and promises this one stays within ‘rules of engagement’ — The Register (Jessica Lyons), 2026-11-05 [blog]
  29. OpenAI launches GPT-6 Cyber with ‘Critical’ cybersecurity rating — Reuters via The Indian Express, 2026-11-06 [blog]
  30. OpenAI released a new cybersecurity model. Is that enough? — TechCrunch (Tim Fernholz), 2026-11-05 [blog]
  31. Threat actors use agentic AI tooling to find and abuse exposed secrets — Microsoft Security Blog (Microsoft Threat Intelligence, Microsoft Security Research), 2026-11-05 [blog]
  32. Rogue AI tasked with renumbering pages accidentally deletes law professor’s entire syllabus, one year of lecture notes — Tom’s Hardware (Stephen Warwick), 2026-11-06 [blog]
  33. Palo Alto Networks in talks to buy AI agent security startup Clutch for $500M — TechCrunch (Marina Temkin), 2026-11-05 [blog]
  34. More visibility and control over Claude Enterprise for security teams — Claude, 2026-11-05 [blog]
  35. AI agents are a ‘bane’ for some website operators, says Cloudflare — The Register (Simon Sharwood), 2026-11-06 [blog]
  36. Introducing Personal Finance in Claude — Claude, 2026-11-05 [blog]
  37. Claude AI can plug directly into your bank account — The Verge (Jess Weatherbed), 2026-11-05 [blog]
  38. Introducing Inkling 2 Small — Thinking Machines Lab, 2026-11-05 [blog]
  39. Tencent Launches Hy4, Joining the Ranks of the World’s Leading Open-Source AI Models — Tencent, 2026-11-04 [blog]
  40. Perplexity launches its own take on AI agents for work — The Verge (Emma Roth), 2026-11-05 [blog]
  41. Airbnb says it’s launching an AI travel agent in the first half of 2027 — TechCrunch (Sarah Perez), 2026-11-05 [blog]
  42. Meta says US users can now video chat with their Muse AI avatar — TechCrunch (Sarah Perez), 2026-11-05 [blog]
  43. Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips — Google (Anish Nangia), 2026-11-05 [blog]
  44. AWS Kiro aims to solve identity crisis by adding another identity: bot orchestrator — The Register (Thomas Claburn), 2026-11-05 [blog]
  45. Copilot Chat has usage caps, and Microsoft would rather you found them by hitting them — The Register (Richard Speed), 2026-11-05 [blog]
  46. Box taps Nvidia’s Nemotron for new agent-first content retrieval system — TechCrunch (Russell Brandom), 2026-11-05 [blog]
  47. Dropbox acquires Tigris Data for $350M to bolster its AI infrastructure — TechCrunch (Lauren Forristal), 2026-11-05 [blog]
  48. Google is ditching Firebase Studio soon, and it wants you to use its newer AI tools instead — Engadget (Ian Carlos Campbell), 2026-11-05 [blog]
  49. Tesla unveils AI accelerator roadmap, adopts new ‘SI’ branding in ‘Super Intelligence’ push — Tom’s Hardware (Anton Shilov), 2026-11-06 [blog]
  50. Underdog shows off a 32MB speech model in new demo — TechCrunch (Ivan Mehta), 2026-11-06 [blog]
  51. How scientists use Claude — Anthropic (Johannes Hermle, Ben Jones, Lukas Althoff et al.), 2026-11-05 [blog]
  52. The Stacks Project, formalized in Lean — Axiom Math, 2026-11-05 [blog]
  53. Axiom Math wants to keep AI slop out of mathematics — TechCrunch (Aditya Mehta), 2026-11-05 [blog]
  54. New NVIDIA Cosmos 3 Open Models for Interactive World Simulation — Hugging Face (NVIDIA), 2026-11-05 [blog]
  55. How to build an agent skill optimizer with AlphaEvolve and Harbor — Google Cloud (Ivan Nardini, Anant Nawalgaria), 2026-11-06 [blog]
  56. Egyptian and Nigerian English join the Open ASR Leaderboard — Hugging Face (Eric Bezzam, Ashley Mo, Aoi Otani), 2026-11-06 [blog]
  57. Which AI is the best liar? — Dwarkesh Patel, 2026-11-05 [video]