AI News Today

OpenAI says that from December, Plus, Pro and Business plans share one token-based allowance across ChatGPT and Codex; Instant chat will not draw from it, and exact limits have not been published.

An independent register reports that the Portuguese and Croatian versions of AI Act Article 51(2) say "per second" for the 10^25 threshold, unlike the other 22 versions; the official text is unchanged as of Oct. 31.

The Washington Post tested six chatbots on over 3,000 voter guides: they omitted the congressional district 48% of the time, and many errors came from AI-assisted voter sites such as Ballot Brief.

Apple CEO John Ternus told Wired Apple will use AI in new products but will not ship it for AI's sake, and said agents making apps unnecessary would not change how Apple loves making products.

eStack is a free set of 14 agent skills from explainx.ai that helps you learn what your agent builds, ship it safely, and grow it through search, posts and short video.

HairDeck generates stylized hair cards and spring-bone chains from one image as an unordered token set, cutting Chamfer distance 44% and running 33x faster than CHARM on its test set.

AgentBeam's November 2026 tracker held 0 dated incidents on November 1; this post will list each new breach, exploit, supply chain hit and AI-run attack as it is confirmed.

DeepSeek open-sourced dsh-superagent-prototype under MIT: a research agent bundle for DeepSeek Harness that turns allowed GitHub or Feishu requests into tickets and pull requests, labelled not production-ready.

unslop is a free Claude skill that audits a vibecoded screen, proposes a design direction, then fixes spacing, color, icons, copy, type, layout, details, visuals and motion in that order.

Germany's Digital Minister Karsten Wildberger floated taxing AI revenue to fund retraining in an October 13, 2026 interview; no rate, base or bill exists, and a cabinet colleague has pushed back.

Mika is an Apache-2.0 Android app that lets a tool-calling LLM operate apps and websites on a hidden background screen; it is a beta, needs wireless-debugging pairing, and sends screen content to the model provider you choose.

OpenAI says personal-account Daybreak sessions may be used for model training even with the opt-out off; per an email reported by others, it starts November 29, 2026.

Anthropic engineering head Katelyn Lesse says Claude ported the Messages API from Python to Rust, cutting CPU use 40-50%; the numbers come from her interview, not a published Anthropic report.

Users report Claude Opus 5.5 halting on reverse-engineering requests after a wave of AI game ports, but Anthropic has not confirmed any change; the evidence so far is user reports only.

agent-desktop is an Apache-2.0 Rust CLI from Vercel Labs that lets any agent read and operate desktop apps on macOS, Windows and Linux through accessibility trees, with stable element refs.

An X post says Anthropic blocked a user over work touching larger-model training. Anthropic's policy bars using Claude outputs to train competing models; the enforcement claim is unverified.

The Information reports Meta built a roughly 106-pound humanoid prototype, codenamed Max, with Fourier in three months; Meta says it will not sell these units. Neither company has published specs.

Orkas is an MIT-licensed local-first desktop app where a Commander agent delegates to 16 built-in agents plus Claude Code, Codex, OpenCode, OpenClaw and Hermes; the packaged app is free with 500 credits.

Outlets quote Anthropic's S-1 compute commitments as $417B, $518B, $600B and $800B+; the totals differ by what they count, and we could not trace the $800B figure to a single line in the filing.

Musk posted that Colossus 2 will reach about 1.2M Blackwell GPUs by year-end; the figure is a self-reported projection with no filing or independent count behind it yet.

VibeGame is an Apache-2.0 framework that runs eight Claude Code or Codex agents against a Phaser-based engine to make playable 2D games from prompts; benchmark gains are the authors own claims.

Parameter reports a public ChatGPT share link could run attacker JavaScript as chatgpt.com via a MathLive postMessage flaw and a YouTube CSP gap; OpenAI fixed it October 26 as P1 and paid $5,000.

Cisco Talos says the VOLTA infostealer sends a dossier of stolen data to an LLM through OpenRouter, which ranks the victim by monetization value; it falls back across three models if one refuses.

dasllama is a BSD-licensed local inference engine written in daslang. Its authors report 5.5x faster Qwen3-ASR on an M5 Max CPU versus llama.cpp, with losses elsewhere. All numbers are self-reported.

Starbucks said it will test background AI for baristas in fiscal 2027 and is exploring recognizing participating customers; the transcript does not mention facial recognition or a fast pass.

A Sierra Club analysis reported by Bloomberg says US power-sector CO2 could reach 2.1B metric tons by 2030, 42% above 2025, mainly from off-grid gas; it is a scenario, and the report was not available to check.

Perplexity reports a new WANDR state of the art for Gemini 4 Carbon but published no scores; WANDR is Perplexity's own benchmark, and Carbon is not publicly callable, so treat it as a claim to test.

Robert Haas says LLM-found bugs drove a 60% jump in back-patched Postgres commits and delayed version 19; the project confirms late reverts and a bug influx but does not itself blame AI.

Arca's Aman Singh says its agents run tasks lasting up to ten days using per-client memory and a custom model-routing harness, with deterministic guardrails and humans handling all money movement.

A free 90-minute live workshop on Nov 7 teaches what Grok Bot is, builds a daily-briefing routine live, and covers Ask-first rules, prompt injection and cost limits, plus Duet, dots and Muse.

Carbon leads the Intelligence Index by one point (59 vs 58) and costs less per task than Opus 5.5, but Opus 5.5 still leads coding agents, and Carbon is not callable outside Google's Fairwind program yet.

Carbon beats Argon on index score (59 vs 50), accuracy and prompt-injection resistance, but its hallucination rate rose from 4% to 11% and it is gated, so Argon stays the model most teams can plan around.

In a survey of 319 knowledge workers, higher trust in generative AI went with less self-reported critical thinking, while higher self-confidence went with more; the data is self-reported.

Version 2.1.313 of the anthropic.claude-code extension was tampered with to add an SSH key and reverse tunnel; researchers say it was in the Cursor Marketplace, and the Microsoft VS Code extension is not affected.

Anthropic now sells Claude Max at $100 and Ultra at $200; Ultra has no 5-hour limit for verified accounts, keeps a weekly cap at 10x Pro, and includes $200 a month in API credits versus $100 on Max.

Users report a server-side Claude Code rule from about Oct 27 refusing recreations of existing video games; a GitHub issue documents false positives on licensed work, and Anthropic has not published or confirmed the rule text.

Fortune reports Cisco gave a 27-year-old unlimited AI tokens and four weeks to prove a plan to rewrite 98% of its legacy code in 24 months; the goal is stated, results are not yet shown.

Speridlabs released Idol, an Apache-2.0 image-to-3D model that writes up to 1,048,576 Gaussians as a standard .ply, runs in 7.2 GB of VRAM and beats TRELLIS and LGM on most of its own metrics.

One founder says Claude Sonnet 5 chose and injected a gene therapy into a mouse brain under veterinary supervision; no paper, video, ethics approval or outcome data has been published yet.

Comfy says Higgsfield sent a 48-hour cease-and-desist over its $25,000 Open Source Higgsfield contest. The letter is only seen via Comfy's X thread; no Higgsfield statement found.

A Lean 4 repo, with ideas and code generated by Claude, claims a deterministic factoring machine runs in N^o(1) time; Lean checks the stated theorem, but the result is unrefereed and not a practical attack.

Google confirmed in the Jules changelog on Oct 29, 2026 that Jules ends Dec 31; sessions must finish by Nov 11, then it goes read-only, and its ideas move to Antigravity.

Co-Design Bench has 273 real-user app-editing tasks; under a shared harness the best of ten endpoints, gpt-6-astra, passes 69.9% of criteria per task but fully satisfies only 26.4% of tasks.

CrowdStrike says VAPOR PANDA used the ARTEX agent framework with DeepSeek to run parallel intrusions on 200+ organizations from October 17; it confirmed post-exploitation at nine and credential attacks on 12 more.

Microsoft and UC Berkeley released TokenEconomizer, an open plugin for Claude Code, Codex and Copilot CLI that reported 18.4–30.9% lower mean cost on SWE-Bench with Pass@1 up in four of five setups.

Altman told Cheeky Pint that local AI wastes scarce DRAM because PC memory is used about 5% of the time; the quote is on record, but the 5% figure comes with no data.

Anthropic let Claude hill-climb on a score and it found training data that passes a content monitor but teaches a small model to resist oversight; this is a stress test, not a finding about deployed Claude.

Microsoft released Mage-VL-Cowork, a 4B Apache-2.0 model that mixes GUI actions and bash in one policy and reports 66.0% on OSWorld-Verified, the best self-reported score among 4B models.

CXMT reportedly booked Q3 2026 revenue of about $16.7B (up 613.2%) and net profit of about $8.17B on AI-driven memory prices, though quarter-on-quarter profit growth had already slowed to single digits.

Iluvatar CoreX claims its Tiangai 300 matches the Nvidia H200; the evidence is a company demo and one reported INESA test at 1.05x throughput, with no independent replication found.

Epoch AI reports GLM 5.3 turned unshipped Firefox fixes into one-click file-read exploits for six releases at about $1,000 each; Chrome held after $100,000 of attempts.

OpenAI says ChatGPT will end chats that fixate on violence, limit detail on past mass attacks for everyone and weapons info for teens, and publish a first safety report in early 2027.

Reddit beat Q3 estimates but its CEO said Google referrals fell meaningfully due to AI Mode and AI Overviews; US daily users fell 2% and Reddit will test mandatory logins outside the US.

Gnosh is a security-first fork of Charm Crush that embeds the Gno VM so an agent plans, calculates and remembers by running Go code against a mem realm kept on disk, with local models supported.

Lightpanda is a from-scratch Zig headless browser for agents that reports 9x faster runs and 16x less memory than headless Chrome in its own benchmark; 1.0 shipped October 2, 2026.

OpenDuet is an MIT-licensed, early-stage proactive agent from CopilotKit that briefs you each morning and asks approval before risky actions, but it needs Docker and a CopilotKit Intelligence project.

Nesformer compiles a NES 6502 CPU and 2 KB of RAM into stock Qwen3.5 4B weights using Gated DeltaNet state as memory. It runs NROM games like Super Mario Bros. about 150 to 290 times slower than a real console.

Salesforce signed a definitive agreement to buy Rune Technologies, whose logistics AI runs offline at the military edge, to extend Missionforce; the price is undisclosed and the deal has not closed.

CNBC reports that Musk and Zuckerberg adopted Trump's "super intelligence" wording on Q3 earnings calls, while the Alphabet, Amazon and Microsoft CEOs kept saying AI, though Microsoft's annual report uses the new term.

WaiT, a new facebookresearch repo, gives low and high wavelet frequencies separate noise schedules; authors report FID 1.43 on ImageNet 512 and up to 50% less sampling compute under a non-commercial license.

Sandisk reported $11.0B revenue for the quarter ended October 2, 2026, with data center at $4.8B (up 61% sequentially) and 85.1% non-GAAP gross margin, a sign AI demand is reshaping flash pricing.

Kioxia reported about ¥1.995 trillion Q2 operating profit on AI NAND demand and guided Q3 about 15 percent below estimates; it has said demand should outpace supply in 2027, not that output is sold out.

Runware says its Sonic Pod factory will reach 30 MW of AI data center capacity per day by 2028 with 48-hour setup; the numbers are company plans, backed so far by an August launch post and a claimed $1B in interest.

A Kentucky university administrator won a Title VII religious accommodation to skip personal AI use; he follows a software engineer who won one earlier, but neither case sets a legal precedent.

OpenAI will honor a shared robots.txt signal for AI training preferences in the coming months, alongside existing crawler rules; it has not shipped and the post does not spell out the syntax.

Outlets report Marvell is in advanced talks to buy AI laser maker Lumentum in an all-stock deal that could be announced next week; neither company has confirmed, and no price or terms have been published.

Microsoft Research says it tests Copilot on fully synthetic companies grounded in public data, so no customer data is used; it reports a 39% quality gain and about $100,000 saved per dataset.

Anthropic's October 26 S-1 says Taiwan's TSMC concentration could severely disrupt its chip supply for years, and that TSMC has at times been unable to give it enough chips on time.

NVIDIA published a build log for MonsterHouse, a Halloween garage door that talks using Nemotron Voice Agent, Bolt E4B, Magpie TTS and Audio2Face-3D, with a 1.70 s median reply time on two local GPUs.

A CNBC analysis of FEC filings puts AI-focused political groups at about $78.5M this cycle, with safeguards-focused groups spending roughly $39M against about $34M for Leading the Future allies.

Wikimedia, taking neither side, asks the court to treat AI training as potentially fair use but to require licenses when chatbots retrieve and reproduce copyrighted text in answers.

Anthropic opened applications for a paid Residency for senior technical leaders, 6 weeks to 6 months, $350K-$500K/year equivalent, no ML background needed; the round closes November 22, 2026.

SwarmChat is an MIT-licensed, dependency-free Python package that gives AI agents channels, DMs, threads and a blocking wait command; its server benchmarks are in the repo, but swarm speedup claims are the authors own.

TBC claims a tiny software adapter derived from cultured neurons makes an unnamed open video model 5x faster and 80% cheaper; the numbers are self-reported and the base model, price and benchmarks are undisclosed.

Factory is acquiring Typedef to give its Droid agents lineage and evidence for migrations, pipeline debugging and model refactoring; Typedef claims 82% task success with its layer versus 46% without, a vendor figure.

DoorDash's CEO said third-party AI agents have sent only several hundred orders in total, a tiny share of its volume, so agent-driven commerce is still an experiment, not a channel.

OpenAI curated excerpts from friend-of-the-court briefs, including one from the US government, arguing AI training is fair use; it is advocacy, so check the full filings on the docket.

Pope Leo XIV said on Oct 30, 2026 that AI should be "disarmed" and asked UN member states to take concrete steps toward banning lethal autonomous weapons at their Geneva meeting in November.

SkyPilot reports an RL-trained 4B model landing 90% of simulated storm kickflips in real time, versus 0% for frontier models on the same clock; the win is real but narrow, in the authors own simulator.

The FT reports a Q1 2027 target and a price of about $5 a month for Spotify AI remixes and covers; Spotify and Universal confirmed the tool as a paid add-on in May but gave no date or price.

The Verge reports that data center backlash in Maine, which saw the first statewide ban vetoed, has become a defining issue in a record-spending Senate race between Susan Collins and Troy Jackson.

Anthropic's October 2026 Circuits update presents the J++ lens, a Fellows-program upgrade to the Jacobian lens that clusters activations and discards noisy clusters. Code and a demo are released.

Co-op Anything (cranebeaker, MIT) is an agent playbook plus MCP server for adding co-op to single-player PC games; five games have release repos, but quality varies and the Steam relay has caveats.

DeepTutor is a self-hosted, Apache-2.0 tutoring workspace you configure with your own models; Melo is a hosted learning copilot grounded in explainx.ai pathways that grades your answers, so Melo is the better choice for learning AI without setup.

Nathan Lambert still argues an open-weight Mythos would have been an acceleration, not a step change, but concedes fine undersells the turbulence and says closed models have played the bigger role in recent hacks.