The DeepSeek Harness team has written a long article on X about designing an agent that stays online, keeps an eye on chats and code platforms on its own, and moves several tasks at once. It describes architectural ideas rather than shipped features, but plenty of it can be borrowed today.
xAI's docs (branded SpaceXSI) say text from grok-4.7, 4.6, 4.5, 4.3 and three grok-4.20 models will carry an invisible watermark by default from Nov 7, including structured outputs and tool-call arguments. It shapes token choice, adds no characters and is flagged by an x-watermark: true header. Admins can opt out per model only if a legal exception applies.
Why it matters: A leading lab is turning on text watermarks by default to meet the EU AI Act's marking duty, a sign watermarking is becoming a baseline for offering models in the EU.
The two sentences Anthropic added to Claude Code cover one client. Anthropic says they target Opus 5.5, but XDA found a second delivery channel that applies them to every model on an account, and Valve says it doesn't want to gatekeep developers' tools. The cost of AI cloning is landing on indie studios that are pulling public playtests and writing "no AI analysis" into their tester agreements.
OpenAI says that from December, as Work's agent features move into ChatGPT, Plus, Pro and Business plans will share one token-based allowance across ChatGPT and Codex. Usage depends on model and task; Instant chat won't count. Combined usage and reset times show in Settings > Usage, and users can wait, upgrade or buy credits at the limit.
Why it matters: Pooling ChatGPT and Codex into one token-based allowance means agentic and coding tasks beyond everyday chat will draw directly on the same plan allowance, changing how paying users use their plans.
Browser UseOfficial@browser_use
Released: BU Ultrafast is live for everyone⚡
Fastest and cheapest web agent on the planet.
> tasks in under 3s
> median task takes 14.4s
> $0.126 total for 25 tasks in the video
Real time, not sped up.
Try it on our cloud today 🔥A federal court in Northern Mississippi on Oct. 30 (US time) denied xAI's motion to dismiss the NAACP's Clean Air Act suit over unpermitted gas turbines in Southaven that power its Colossus data centers. xAI, now owned by SpaceX, argued the turbines were exempt 'mobile sources' and citizen suits are unconstitutional; the judge rejected both. It rejected a DOJ dismissal bid about two weeks earlier.
Why it matters: The case now moves forward on the question of whether AI data centers may run on-site gas turbines without air permits, an approach several AI companies use to add power quickly.
PlanAssist, a Gemini-based tool from Google DeepMind, the UK government, Google Cloud and Faculty, helps officers assess home planning applications. Pilots in Barnet, Camden and Dorset cut average assessment time 58%. A beta of 12 months adds 12 councils (15 total); if tests succeed, the government plans to offer it to all councils in England in 2027, with a qualified officer approving each one.
Why it matters: It is a case of a frontier lab's model entering a national public approvals process, with measurable pilot results and a human keeping the final say, a reference point for other governments adopting AI.
OpenAI released GPT-Rosalind-6, a reasoning model for biology, drug discovery and translational medicine, offered as GPT-Rosalind-6 Research via trusted access to approved scientists, institutes, firms and government partners. It shares weights with GPT-6 Astra and refuses fewer dual-use biology queries but still refuses bioweaponization. It runs in Work and Codex; the API needs separate approval.
Why it matters: OpenAI is offering a flagship-level model with looser dual-use biology limits only to vetted organizations, a precedent for tiered access to frontier models in high-risk scientific fields.
DeepSeek open-sourced dsh-superagent (MIT), a prototype coding agent on DeepSeek Harness for one GitHub repo, optionally Feishu. A main session vets requests, ticketing approved ones; each runs in its own worker session and folder, replying in the source thread, e.g. a PR; a web panel lists what needs a person. DeepSeek calls it research code, not a product; parts of its article are unimplemented.
Why it matters: DeepSeek published a working always-on coding agent that takes work under team rules, runs each ticket in isolation and delivers pull requests, an open reference for teams deploying autonomous coding agents, though still a prototype.
Talos found VOLTA-AUTODOXER, a self-deleting Windows infostealer. After stealing passwords, cookies and crypto wallets, it uses stolen sessions to mine email receipts and food-delivery orders, then sends the dossier via OpenRouter to Claude Opus 4.6 for a target assessment, falling back to Kimi K2 and Gemini 2.5 Flash on refusal. Talos has not confirmed use in the wild.
Why it matters: Malware is now using LLMs to do the attacker's analysis, with a built-in chain to switch models after refusals, testing model guardrails in a new way.
The robots.txt token backed by Google, Microsoft, OpenAI and Anthropic standardizes one word: training. The same two lines touch only training at Microsoft, but at Google they also cost a site its place in Gemini's cited answers. The choice small sites care about, cite me but don't train on me, still has to be worked out company by company.
Microsoft's Mage team released Mage-VL-Cowork weights on Hugging Face under Apache-2.0. The 4B-parameter vision-language model operates both the GUI (screenshots, mouse, keyboard) and code/terminal within a single policy. Microsoft says it scores 66.0% on OSWorld-Verified with a 50-step budget, above larger models such as Qwen3.6-27B; online RL adds 10.16 points over hybrid SFT.
Why it matters: A small, commercially usable open model approaching larger agents on computer use makes on-device agents more practical.
Anthropic added a server-side rule to Claude Code on Opus 5.5, with no off switch: original, genre and own-asset games are fine, but not a recreation of a specific existing game, even renamed. A spokesperson pointed to staffer Addy Osmani, who said using Claude to infringe IP breaks its usage policy. XDA found the instruction in Claude Code v2.1.314.
Why it matters: It is a server-side rule users cannot switch off that stops a major AI coding tool from recreating specific existing games. Anthropic cites its usage policy on IP infringement, a sign model makers are handling copyright risk in-product.
OpenAI's Oct 30 ChatGPT release notes say account switching is rolling out to the iOS and Android apps, after launching on the web. Users can stay signed in to up to two ChatGPT accounts and switch via Settings > Switch account without logging out. Chats, memory, files, settings and subscriptions stay separate per account, and admins of managed workspaces can control availability.
Why it matters: On mobile, work and personal ChatGPT accounts can now stay signed in together with no logging out, while chats, memory and subscriptions stay separate and managed-workspace admins control access.
S&P Global Ratings on Oct 30 moved Meta's AA- outlook to negative from stable, citing uncertainty over its capex and free operating cash flow and saying Meta's ability to monetize AI is more limited than peers'. S&P sees FOCF deficits of $39B in 2026 and $71B in 2027. Meta guides $135B-$145B capex this year; a spokesperson said it respectfully disagrees with S&P.
Why it matters: S&P says Meta's ability to monetize AI lags peers' and projects negative free operating cash flow in 2026 and 2027, a rating-agency signal on the credit risk of hyperscale AI capex. Meta disagrees.
OpenAI has secured about $11 billion in commitments for its latest round, under half its initial $30 billion target, and expects to raise more in coming weeks, the WSJ reported, citing people familiar with the matter. The FT reported OpenAI cut its target to about $20 billion and its valuation goal from $1.4 trillion to $1.25 trillion as some investors trimmed commitments amid an AI selloff.
Why it matters: OpenAI reportedly has commitments for under half its $30B target, with its valuation goal cut to $1.25T from $1.4T. Amid an AI selloff, it suggests investors are more cautious on the top-valued AI firm. OpenAI hasn't announced it.
OpenCode got Claude subscriptions back not by holding users' logins but by running the copy of Claude Code they already installed. It works, but which subscriptions an open-source coding agent can use, and how they're billed, is still set by the lab's own client and help pages.
Dulo, founded by Google X founder Sebastian Thrun and Adrian Aoun, has raised a $130 million seed round, first reported by The Information and confirmed by Aoun on LinkedIn. Backers include Eric Schmidt, John Doerr, Marc Benioff, Yuri Milner, Construct Capital and Lowercarbon Capital.
Why it matters: It is an unusually large seed round in AI-for-engineering, a field that also includes the Bezos-backed Prometheus.
We’ve released AA-Omniscience v1.1, updating our knowledge and reasoning benchmark with a refreshed public set, removal of saturated questions, and ambiguity fixes
AA-Omniscience is our proprietary benchmark measuring model knowledge and hallucinations, including whether models admit when they don’t know an answer. Our primary metric, Net Score, rewards correct answers, penalizes incorrect answers, and treats abstentions neutrally.
AA-Omniscience was updated to v1.1 as part of the Artificial Analysis Intelligence Index v5 release.
Key changes in AA-Omniscience v1.1:
➤ Updated 3,000-question dataset: Down from 6,000 questions, with previously public and saturated questions removed, errors and ambiguities corrected, and new questions added
➤ Newly calibrated 300-question public set: AA-Omniscience v1.1 has a new public set on @huggingface calibrated to match the difficulty of the private set. At release, public and private scores have a 0.99 Pearson correlation and 0.99 Spearman rank correlation for AA-Omniscience Net Score, with a median difference of 0.0 points. Accuracy and hallucination rate also have 0.99 correlations, with median differences of -0.07 and -0.29 percentage points, respectively
➤ Slightly updated domain distribution: Questions span the same 6 domains and 42 topics as version 1, with a revised distribution: Science, Engineering and Mathematics (600), Software Engineering (600), Finance (450), Health (450), Humanities and Social Sciences (450), and Law (450)
Key results for AA-Omniscience v1.1:
➤ Rank order largely preserved: AA-Omniscience v1.1 has a 0.98 Spearman rank correlation with v1. The recently released Gemini 4 Carbon (high) leads with a Net Score of 54, followed by Claude Opus 5.5 (48, max with fallback) and Claude Fable 5.1 (46, max with fallback). Muse Spark Open (xhigh) is the leading open weights model, 28 points behind Gemini 4 Carbon
➤ Scores average ~2 points lower on the refreshed set: Models score 1.9 points lower on Net Score on average compared to v1, with a median decrease of 0.6 points. Some models see much larger declines, particularly MiniMax M3.1 Flash, down 22 points (16 to -6), and MiMo V2.6 Pro, down 13 points (8 to -5)
➤ Slightly lower accuracy: AA-Omniscience Accuracy measures the percentage of questions the model answers correctly. Models answer a median of 1.5 percentage points fewer questions correctly on average compared to v1. Claude Fable 5.1 (max with fallback) has the highest accuracy at 68%, followed by Claude Opus 5.5 (max with fallback) at 66%
➤ Hallucination rates decrease: AA-Hallucination measures how often models choose to answer when they do not know the answer. The median Hallucination Rate falls 1.4 percentage points, with Gemini 4 Argon (high) recording the lowest hallucination rate at 4%On its Oct 30 third-quarter earnings call, MediaTek said it has won a second customer for its custom data-center AI accelerators. It did not name the customer but said it is not Chinese, after speculation about ByteDance; production and first revenue are expected in 2028. CEO Rick Tsai said the second-generation chip for its existing customer is on track for volume production in Q1 2028.
Why it matters: MediaTek's custom AI chip business grows from one customer to two; the company is aiming for the high end of a roughly $16 billion revenue range for it in 2027.
OpenCode v2.2.3, released Oct 30, adds a built-in Claude Agent provider: with Claude Code installed locally and `claude auth login` run, users select Claude Agent in /connect to use their Claude Code login and models. In March 2026, v1.3.0 removed OpenCode's built-in Anthropic OAuth login plugin following legal requests from Anthropic.
Why it matters: OpenCode v2.2.3 adds a built-in Claude Agent provider: after installing Claude Code and logging in, users can use their Claude Code login and its models inside OpenCode.
Ludo Robotics, the Palo Alto subsidiary of Krafton, presented Ludi 0.2 at UC Berkeley's 2026 Bay Area Robotics Symposium, improving intent reading, memory and reasoning, gesture recognition and movement coordination. Ludi 0.1, introduced in August, paired a fine-tuned vision-language model with robot-control policies. A unified foundation model, Ludi 1.0, is in development.
Why it matters: A major game publisher is carrying its AI-teammate work into physical robots, aiming at intent understanding, one of the key hurdles for robots that work alongside people.
Washington-based chip startup Powerlattice has sold $86 million of equity toward a $111 million offering with 18 investors, an SEC Form D filing shows; investors and round are not named. Its power delivery chiplet sits inside the processor package and, the company says, cuts AI compute power needs by more than 50%. Former Intel CEO Pat Gelsinger is on its board; it had raised $31 million before.
Why it matters: Power and cooling are becoming the main limits on AI data center growth; a raise nearly triple its prior funding shows investors backing ways to cut accelerator power use at the chip-package level.
Apple said the first Apple Intelligence features, including Live Translation, Writing Tools, Visual Intelligence, Clean Up and Genmoji, are coming to iPhone users in China. Via an Alibaba partnership, Siri, Writing Tools and Visual Intelligence will also support Qwen. Testing starts in the iOS developer beta on November 4, with release in iOS 27.3 in December.
Why it matters: Apple Intelligence had been unavailable in China; launching it with Apple's own models plus Alibaba's Qwen brings it to one of the largest smartphone markets and shows how foreign tech firms can ship generative AI under Chinese rules.
OpenAI added ChatGPT safeguards: users predicted to be under 18 get tighter limits on practical weapons and security-system information even for harmless-looking questions; for all users, ChatGPT gives less detail on past mass attacks and can end chats fixated on violence, pausing repeat accounts. Reuters says rollout takes the next few weeks, with parents notified of teen pauses.
Why it matters: It tightens how a leading chatbot handles teen safety and violence prevention as AI firms face growing scrutiny over risks to minors, and may shape industry content-safety norms.
Google will support a shared AI training control in robots.txt based on an IETF AIPREF draft: one 'policy token' states a site's training preference to many AI providers at once, while controls like Google-Extended keep working. The Verge reports OpenAI, Microsoft and Anthropic will honor it too; Anthropic already does, Google and Microsoft plan support next year, OpenAI in coming months.
Why it matters: It is the first time the major AI labs have lined up behind one opt-out signal for training, simplifying crawler control for publishers and shaping how training data is gathered.
Anthropic renamed its top Claude plans: the $100/month Max 5x is now Max and the $200/month Max 20x is now Ultra, prices unchanged, subscribers moved automatically. Verified Ultra accounts lose the rolling five-hour limit, though the weekly cap (10x Pro) stays; new accounts keep a temporary limit until verified. The plans include monthly API credits: $100 on Max, $200 on Ultra.
Why it matters: The five-hour cap has been a major pain point for heavy Claude and Claude Code users; removing it changes how long-running coding and agent workflows can be run on a subscription.
Google Research says its Project Suncatcher prototype satellite, launched earlier this month, finished commissioning and ran early ML tests on its TPUs. This week the onboard Gemma 4 open model answered an uplinked prompt with a space haiku, which Google calls the first response from a data center class TPU in orbit. Thermal and radiation tests are next; twin satellites are planned for 2027.
Why it matters: It is the first in-orbit check of Google's plan to power AI compute with solar energy in space, one possible route around the power and cooling limits facing ground data centers.
Amazon's latest 10-Q shows a secured loan facility of up to about $80 billion it can draw from Q1 2027 through Q3 2028 to buy technology infrastructure. Borrowings carry variable rates, mature five years after each draw and are secured by some purchased assets. Amazon still expects about $220 billion of 2026 capex, rising meaningfully in 2027, mostly on tech infrastructure.
Why it matters: It is one of the largest dedicated borrowing lines a hyperscaler has set up for AI infrastructure, showing AI buildouts are leaning more on debt, a backdrop to current market worries about AI spending.
Salesforce has signed a definitive agreement to acquire Rune Technologies; the deal has not closed. Founded in 2024, Rune makes logistics software for frontline units that works without reliable connectivity and is deployed with the US Army and Marine Corps. It will join Missionforce, Salesforce's defense unit, in what the company calls its most significant national security deal.
Why it matters: It shows a major enterprise software company pushing AI agents into frontline defense work, deepening Salesforce's bet on the military AI market.
The EU's DMA acquisitions list shows Apple struck a deal with Ascension Labs (Ascend.io) to hire certain employees, take a non-exclusive license to its IP and be assigned some patents. Ascend.io uses AI agents to automate data pipeline ingest, transformation and orchestration; its site says it is winding down. Apple had notified the European Commission of the deal earlier.
Why it matters: It is another of Apple's small hire-and-license AI deals, pointing to investment in internal AI and data infrastructure, and shows how DMA filings are making such big-tech deals public.
The UK CMA opened a Competition Act 1998 case into whether Meta abuses a dominant position with terms that could let it take content from publishers' and businesses' websites for its own products, including AI features, without giving them a choice. Meta had already paused the terms and said there were "no grounds" for the probe. The CMA says no breach has been found yet.
Why it matters: Regulators are now formally testing whether platforms can use business terms to source AI training content, which could shape how big platforms obtain data for AI.
Cursor said a tampered Anthropic.claude-code 2.1.313, not built by Anthropic, was on its extension marketplace for about 11 hours on Oct 29. It collected host details, added an attacker SSH key and tried to open a persistent reverse tunnel, though enabling SSH fails on most Macs and Windows PCs without admin rights. Up to 5,000 users were affected; Cursor pulled it and published cleanup scripts.
Why it matters: AI coding extensions run directly on developer machines, and this shows extension marketplaces are now a supply-chain entry point for attacks on developers.
Fixed PHP images went public six hours before the agreed time, and the open-weight model GLM 5.3 used Firefox fixes that hadn't yet shipped to break six releases. Both point to one shift: as bugs get cheaper to find, the gap between a fix going public and users installing it has become open source's security bottleneck. Mozilla has changed how it ships, but its fixes take money, staff and some of the openness open-source projects are used to.
Nvidia is in talks to provide credit guarantees to UK AI cloud provider Nscale to back debt for power and construction at a 460 MW data center on its Monarch campus in West Virginia that Anthropic plans to use, The Information and Bloomberg reported. The Information put the support at billions of dollars. No deal is signed yet. Nvidia already invests in Nscale.
Why it matters: Nvidia would be chip supplier, investor and debt backer in one deal chain, showing how much neocloud financing now leans on its balance sheet.
An internal Sierra Club analysis estimates that if planned AI data centers are built, US power-sector CO2 would hit about 2.1 billion tons in 2030, up 618 million tons or roughly 42% from 2025, mostly from off-grid gas plants. The estimate is higher than others, such as a June academic study projecting up to 28%, and depends on what gets built.
Why it matters: The energy and emissions cost of the AI buildout is becoming central to data center siting, permitting and local opposition, and estimates vary widely.
The Wikimedia Foundation, which runs Wikipedia, filed an amicus brief in The New York Times' copyright suit against OpenAI and Microsoft. It argues training LLMs on copyrighted work should be fair use, while RAG that serves content back to users must give attribution under CC BY-SA 4.0. It warns an overly black-and-white ruling could threaten open knowledge and open-source AI.
Why it matters: The case is a landmark for AI copyright, and Wikipedia is one of the most important sources of training data, so its position could shape how courts separate training from retrieval uses.
Amazon and Intel shares fell 7% and 25% in Friday premarket trading after disappointing September-quarter results, pulling Nasdaq futures down about 2%, CNBC reported. Amazon missed estimates, kept its $220 billion 2026 capex forecast and again posted negative free cash flow. Alphabet fell 13% on Thursday, its worst day on record. The FT says AI spending fears are deepening a global tech rout.
Why it matters: Whether huge AI capex will pay off has become the market's central question, and the slide in big tech shares could affect the pace of future AI infrastructure spending.
CrowdStrike says a group it tracks as VAPOR PANDA, which it links to China, used the open-source ARTEX agent framework against 200+ organizations in 14+ countries in 11 days, confirming access at nine. CrowdStrike says ARTEX ran mainly on DeepSeek plus an unidentified commercial model, used about 1.4 billion input tokens in one intrusion, and a person steered key steps. DeepSeek has not responded.
Why it matters: CrowdStrike says it had mostly seen AI agents assist single intrusions; here one operator ran many in parallel with an open-source framework and publicly available models.
Google says the Gemini app can now generate, preview and edit files inside a chat: users can expand a file to full screen and type edits, highlight text to ask questions, or have Gemini revise it. PDF preview and editing work on web and mobile; Google Docs, Sheets, slide decks, Markdown, TXT and code are web-only for now. It is available to users in every language the Gemini app supports.
Why it matters: Files made in Gemini can now be revised without leaving the app for Docs or Sheets.
