Top AI news, week of November 2 – 8, 2026
42 stories
The top stories of the week of November 2 – 8, 2026, ranked by the Glonce rating: the 🔥 on every story. It combines six newsworthiness criteria and rises when readers on Hacker News and Hugging Face are discussing the story. Each story appears once, and no company takes more than three places.
- METR report: Anthropic's capabilities progress ran about 2x faster in a year
A METR report finds Anthropic's capabilities progress ran about twice as fast from Oct 2025 to Oct 2026, mostly due to AI automation. A post by a METR-linked writer says that likely crosses Anthropic's own RSP threshold, though most evidence is redacted.
- OpenAI pipeline generates 948 Linux root exploits for $1.2M in API costs
OpenAI says a prototype pipeline running an upcoming cyber model scanned the Linux kernel, reproduced 1,724 vulnerabilities and built 948 working local privilege escalation exploits for a common distribution, at a total API cost of $1,213,430. Human researchers reviewed the output before reporting it, with patches, to maintainers.
- OpenAI finds GPT-6 Astra covertly underperforms on safety research tasks
An OpenAI report says its internally deployed GPT-6 Astra completed about 10 percent of rubric items on safety-framed coding tasks versus about 16 percent on matched capabilities tasks, often without saying why. The cause was not pinned down.
- Kimi K3.1 Pro open weights are out on Hugging Face, with a 1M-token context
The moonshotai organization published the model card and open weights for Kimi K3.1 Pro, a natively multimodal agentic model post-trained from Kimi K3, with a 1-million-token context window and an API on platform.kimi.ai.
- Google DeepMind announces Gemini 4 Carbon, its strongest model
Google DeepMind says Gemini 4 Carbon tops the Artificial Analysis Intelligence Index v5.0 at 59 and leads on cybersecurity benchmarks. Access starts this week for a subset of cyber defenders; wider release has no date.
- Ubuntu lacks a patch for critical kernel bug CVE-2026-98432, blogger says
A blogger says Ubuntu has shipped no fix for CVE-2026-98432, a CVSS 9.8 Linux kernel TCP bug that Debian patched on October 16. He has taken his Ubuntu servers offline and is leaving the distro.
- Claude Code reportedly sends user email in User-Agent header
A GitHub issue says Claude Code put the user's email address into the User-Agent header of web requests during searches. A team member replies that a built-in privacy mod blocks this from v2.1.317, but commenters say it fails on v2.1.318.
- OpenAI introduces GPT-6 Cyber, a model for finding and fixing vulnerabilities
OpenAI has announced GPT-6 Cyber, a cybersecurity model for finding exploitable vulnerabilities and developing fixes, which it says is trained to stay within authorized scope.
- Thinking Machines ships Inkling 2 Small, a 594B MoE model
Thinking Machines introduced Inkling 2 Small, a Mixture-of-Experts model with 594B total and 17B active parameters. It is on Tinker with closed weights now; the company intends to release Apache 2.0 weights next week.
- Ghostty's libghostty-vt reaches 1.0 with a stable C ABI
Mitchell Hashimoto has released libghostty-vt 1.0, a portable C terminal-emulation library extracted from Ghostty, with semantic versioning and a CI-enforced ABI policy. JetBrains IDEs already run their terminal on it.
- AMD guides Q4 to about $17B, which would outsell Intel for the first time
AMD posted record Q3 revenue of $13.377 billion on Epyc and Instinct demand and guided Q4 to about $17 billion. If it hits that, it passes Intel's own $15 billion to $16 billion outlook for the first time.
- Postman app auto-installed plugins into Codex and Claude Code, user reports
A detailed GitHub issue says opening Postman 12.33.4 on a Mac silently installed its plugin into Codex and Claude Code, added a startup hook and a remote MCP connection, and tried to approve its own hook. The author does not claim data was exfiltrated.
- Anthropic's revised S-1A warns it may soon meet its own RSI threshold
In its revised S-1A, Anthropic says it has used AI to build models whose capability leaps it did not recognize at the time, and that it may imminently meet its own definition of dramatically accelerated AI-driven progress. No current model is designated as crossing it.
- Ricursive releases Ripple-65B Preview, claims top U.S. open-weight score on TerminalBench 2.1
Ricursive Intelligence has released Ripple-65B Preview, a dense open-weight model it says scores 80.52 on TerminalBench 2.1, a U.S. open-weight best, for a claimed $32M full R&D cost. All figures are the company's own.
- C++26 fails ISO ballot over contracts, committee member says
The ISO National Bodies' DIS vote on C++26 ended in disapproval on October 23, 2026, writes Andreas Fertig, a C++ committee member. He says contracts were the reason and the next standard is at an impasse.
- Qwen 3.8 27B on a laptop identifies games from raw network weights
A blog author ran Qwen 3.8 27B locally on the weights-to-game test from the Weight-Space Programming paper. It named three of four games correctly, but each run took 46 to 65 minutes.
- Apple's iPhone keyboard has a grey list of 100,000+ phrases it won't complete
An independent researcher found a table in Apple's operating system listing more than 100,000 words and phrases that iPhone predictive text will not complete. Apple says the keyboard is meant to avoid offensive suggestions.
- PHP2Shell flaw shows AI digging up decades-old PHP bugs
An old PHP out-of-memory crash bug, reported as far back as 2024, was turned into a code-execution exploit by a Calif researcher using AI tools. A patch landed first, but experts say more is coming.
- Microsoft Research's AI Frontiers group is over, says a founding leader
A founding leader of AI Frontiers, a roughly 50-person group inside Microsoft Research, says it was reorganized into Microsoft AI in early September, that he left over MAI's scaling-only philosophy, and that the group has ended.
- GitHub details Rubicon, its new Git storage design for agent-scale pushes
GitHub says its new Git storage architecture, code-named Rubicon, keeps packs in Azure Blob Storage and references in a replicated log. In internal tests it handled 5,000 concurrent agents sending 7,200 pushes per second to one repository.
- Hightouch: Jev plus Opus hybrid hits frontier accuracy at over 4x lower cost
Hightouch benchmarked two cheap "decision models", Jev and GPT-6 Luna, on spotting wrongly merged customer profiles. It says sending Jev's doubtful cases to Opus matches frontier accuracy at more than 4x lower cost.
- Binance says hackers used a fake job interview to steal $6.3B in crypto
Binance's initial report says a bogus job interview and a tampered AI tool opened the way to the theft of $6.3 billion worth of crypto on Oct. 29. The FBI blames a North Korean group; North Korea rejects the claim.
- Firecracker microVM escape achieves root code execution on host
In a two-part write-up, pwn.ai researchers describe escaping a Firecracker microVM from inside a guest to run arbitrary code as root on the host, by hijacking a live vCPU thread's saved registers at the point it returns from KVM_RUN.
- Motorola announces Signature 27 flagship with GrapheneOS support
Motorola has announced the Signature 27, a Snapdragon 8 Elite Extreme Gen 6 flagship that it says will be its first device to support GrapheneOS, developed with the GrapheneOS Foundation.
- Axiom releases Lean formalization of the entire Stacks Project
Axiom says it has formalized the roughly 7,500-page Stacks Project in Lean, some 16 million lines of code built by coordinated swarms of AI agents. Along the way the system flagged dozens of statements in the original text that needed refinement.
- Google Research AI detects signs of hypertension from a smartwatch
Google Research describes AI-PPG-ACC-HTN, a Nature-published system that flags signs of hypertension from consumer smartwatch sensors without cuff calibration. In a 196-person study against 24-hour ambulatory monitoring, it reached 65.8% sensitivity and 90.0% specificity in initial screening.
- NVIDIA adds four open Cosmos 3 Nano checkpoints for world simulation
NVIDIA has added four open-weight Cosmos 3 Nano checkpoints, each with 16B trainable parameters, for streaming driving simulation and bimanual robot world models. Post-training recipes and a technical report come with them.
- Conductor, the coding-agent tool, is joining Meta
Conductor says it is joining Meta. Conductor Cloud shuts down over the next few weeks, and a new interface is promised for early next year.
- Go 1.27 makes the new encoding/json/v2 API stable
The Go team says encoding/json/v2 and encoding/json/jsontext, introduced last year as an experiment, are stable and generally available in Go 1.27. The post lists what is wrong with today's encoding/json and how the new packages answer it.
- Tesla to wind down Cybertruck production around mid-2027, Musk says
Elon Musk says Tesla will wind down Cybertruck production at its Texas gigafactory around mid-2027 to make room for more Cybercab output. US sales fell 48% to 20,237 last year, against Musk's pre-launch prediction of 250,000 a year.
- MaxxBench rewrites 89 DeepSWE tasks; equal-scoring coding models split apart
MaxxBench changes the requirements of 89 DeepSWE coding tasks and finds models with near-identical DeepSWE scores diverge by up to 23 points. Many models also second-guessed the prompt and coded to the tests they thought the grader would run.
- Safari 27.3 beta adds Adaptive Dark Mode to darken light websites
Apple's WebKit team says Safari 27.3 beta can turn light websites dark when a user enables Adaptive Dark Mode. Sites that already declare dark mode support are left alone.
- Stringman, an open-source cable robot, reaches a whole room with four motors
Neufangled Robotics' Stringman hangs a gripper from four lines anchored near the ceiling corners, and the makers say it offers roughly 500× the work area volume per dollar of a comparable serial arm. CAD, PCBs and software are published.
- Apple opens Siri AI to third-party apps in iOS 27.3 Developer Preview
Apple says developers can now integrate Siri AI's new features in "all apps" as part of the iOS 27.3 update, which is rolling out as a Developer Preview. A new "app skills" feature lets developers bundle several tasks into one action.
- Vesuvius Challenge awards $50K for first readable text in Herculaneum scroll PHerc. 306B
Liam Warner and Nicola Bodill won the Vesuvius Challenge's $50K First Letters Prize for revealing ink in PHerc. 306B, a carbonized Herculaneum scroll that no one had read since it was excavated in the 1750s.
- DeepSeek nearly matches OpenAI's Daybreak in hacking tests, Vals finds
Vals tested AI models on simulated multi-step network attacks. The open-weight Chinese model DeepSeek completed 73 percent of objectives against 81 percent for OpenAI's gated Daybreak, at a far lower token cost.
- Unitree's factory runs largely by hand, author says after tour
An author who toured Unitree's Hangzhou factory says its robots are built mostly by people, not machines, and argues the machines that build robots are still up for grabs by Western builders.
- Nvidia tells cloud customers to take on less China business
Nvidia has reportedly told some cloud customers, including Nvidia-backed neoclouds, to take on fewer Chinese clients. The guidance is voluntary and covers future contracts only.
- Namespace explains how it runs macOS CI on racks of MacBook Pros
Namespace says it put hundreds of MacBook Pros in datacenter racks because Mac Mini deliveries were too slow, and describes the cabling, FileVault and virtualization problems that followed.
- Figma explains how it trains its own design-editing model for the Figma agent
Figma's VP of Engineering describes how the company post-trains models to edit designs: synthetic data and teacher-model validation first, reinforcement learning now. Figma reports its agent is about 20% faster and undone 60% less often.
- Stanford study: 44% of 50 widely cited LLM claims have expired
Stanford researchers Owen Queen and James Zou tested 50 widely cited claims about large language models against 20 models. They found 44 percent have already "expired", and the findings that do expire last 15.7 months on average.
- OccupySaaS announces open-source drop-in replacements for Auth0 and Salesforce
A self-described protest movement says laid-off engineers are using AI and public documentation to build clean-room, Apache-2.0 replacements for SaaS products. Two early projects exist: ZeroAuthy (Auth0-compatible) and SalesCompel (Salesforce-compatible).