AI Morning Brief · Nov 15, 2026

Coverage window: 2026-11-14 07:00 → 2026-11-15 07:00 (Beijing Time)

Research note: English edition written primarily from English-language sources; some items cite Chinese-language sources.

Top Stories

1.FBI issues ARTEX alert: an open-source AI penetration-testing framework has been weaponized, reaching privileged access in under three hours

The FBI issued FLASH-20261113-001 on Nov 13: since August 2026, multiple kinds of malicious actors have used ARTEX (a Go-based multi-agent autonomous penetration-testing framework that can connect to Claude, GPT, DeepSeek, GLM, Grok, Kimi and other models) to automate intrusions across finance, government, defense, semiconductors and other sectors; in one South Korean financial-institution case about 67,000 people's personal data was stolen, and a ConnectWise supply-chain case affected more than 2,250 downstream tenants; the original author deleted the repository in October, but forks and a commercial version, ARTEX PRO, remain in circulation, and more than 1,600 ARTEX consoles were exposed online in early November. (FBI alert via OODA Loop and MLQ, Nov 13; a separate incident from the Unit 42 report on the Cairn framework covered in the previous issue, and the two together show AI-agent intrusions have become a scaled threat)

2.Anthropic IPO follow-up: underwriters reportedly refuse subscriptions from mainland China and Hong Kong investors; retail subscription closes Nov 17

Bloomberg reported on Nov 13, citing people familiar with the matter, that several underwriters are refusing or giving essentially no allocation to institutional investors from mainland China and Hong Kong, citing the US International Traffic in Arms Regulations (ITAR), though the prospectus itself does not state the restriction in so many words; a family office of a Chinese tech founder has applied through a British Virgin Islands trust, with the outcome undecided; the IPO is still set for pricing on Nov 18 and Nasdaq trading on Nov 19 under ticker ANT, and the retail subscription deadline is 16:00 ET on Nov 17. (Bloomberg via TradingView and Stocktwits, Nov 13; single source chain)

3.NVIDIA follow-up: director Mark Stevens sold $946M of stock in Q3 with no 10b5-1 plan; NVDA falls 12.75% for the week, its sixth straight weekly loss, and drops below $4T

SEC Form 4 filings show director and second-largest individual shareholder Stevens sold about 4.237M shares in three tranches from Aug 31 to Sep 18, cashing out about $946M, with no 10b5-1 preset trading plan marked, and still holding about 27.5M shares; Buteon's tally of all Form 4 filings puts total NVIDIA insider selling in Q3 at about $968M, 98% of it Stevens, but similar in scale to prior years. NVDA closed at $162.85 on Nov 13, down 2.01%, for a weekly loss of 12.75%, the longest six-week losing streak since 2012, with the market value falling below $4T for the first time since July 2025; the next test is Q3 earnings on Nov 17, the first call hosted by interim CEO Kress. (Form 4 data via Benzinga and TradingView; weekly figures via MLQ, Nov 13–14)

4.ChatGPT Wallet follow-up: United Airlines bans AI agents from booking tickets and Amazon blocks agent orders, so the first agent shopping hits a wall

Wallet launched on Nov 13 for US Pro subscribers, with OpenAI saying it takes a 0% cut at launch and issuing one-time tokens per payment without exposing the real card number to agents; meanwhile United revised its contract of carriage on Nov 11 (Rule 6.J.6), barring unauthorized AI agents (including OpenAI's dots and Meta's Muse) from buying, changing, seat-selecting or checking in for tickets, with violators risking cancellation, while bookings made before Nov 11 stay valid and fare searches remain allowed with purchases done manually; Amazon is also blocking agent orders. Background: OpenAI acquired AI travel-agent startup Odessia on Nov 10 (terms undisclosed). (AI Agent Update, Nov 14; United's ban was announced Nov 11 and is first recorded in this issue)

5.Doubao Work opens up to third-party models Kimi 3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash

ByteDance's Doubao Work opened third-party models to all users from Nov 12, with the first batch being Kimi 3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash; the model picker shows quota multipliers directly (Kimi 3.1 Pro at 0.95x; Doubao 2.1 Lite and DeepSeek V4.1 Flash at a limited-time 0.01x), and new users logging in on the desktop app get a 30-day free subscription; third-party models were previously available only to enterprise customers. It is the first time models from Kimi, GLM and DeepSeek have appeared together in a ByteDance office product. (Multiple Chinese outlets, Nov 13–14)

6.Google follows OpenAI: Gemini connects to Google Wallet and can analyze real spending and bank accounts

Gemini Personal Intelligence has added a Google Wallet connection: US users can let Gemini analyze spending, track subscription price rises and generate cash-flow charts, and can also link banks and credit cards via Plaid; the access is read-only and cannot move money or invest; it is rolling out to all US Gemini users, not limited to paid tiers. (Neowin, Nov 14)

7.DeepSeek follow-up: pricing details for the new app API are out, with search and web fetch at 5 yuan per 1,000 calls

After the full app-layer API announcement on Nov 13, multiple outlets disclosed pricing on Nov 14: search and web fetch are both 5 yuan per 1,000 calls, OCR is 1.5 yuan per 1,000 pages, ASR is 0.39 yuan per hour at a limited-time 50% discount, and TTS is 1.5 yuan per 10,000 characters at a limited-time 50% discount; only V4.1-Flash (deepseek-flash) supports these built-in capabilities; the open-source frontend demo chat-demo's GitHub repository was actually created on Nov 12, not newly on Nov 14. (Multiple Chinese outlets, Nov 14; this resolves the previous issue's unverified item on the "5 yuan per 1,000 calls" search pricing)

8.The EU holds its first-ever AI summit in Dublin, and Mistral's CEO says "the next two years will decide the next ten"

Leaders of the 27 member states held their first summit devoted to AI on Nov 13, discussing competitiveness, funding disputes over the plan for seven AI gigafactories and sovereignty; Mistral CEO Arthur Mensch and Nobel laureate Geoffrey Hinton briefed the leaders on European competitiveness and the social impact of AI respectively, with Mensch saying Europe lacks speed, funding and coordination rather than technology or talent; the meeting reached no formal conclusions, and the outcomes feed into an informal retreat on Feb 5, 2027. (AI Affairs, Nov 13)

9.CoreWeave seeks at least $4B in investment-grade loans, collateralized by Meta contracts and chips

Bloomberg first reported on Nov 11–12, with several outlets following on Nov 13, that CoreWeave is seeking at least $4B in investment-grade loans, led by Morgan Stanley and MUFG, targeting a rate of about 7–7.5%, with the issue to launch as early as the week of Nov 16, collateralized by Meta contracts and chips; total debt was about $47B at the end of Q3, and the CEO said on the Nov 10 call that the company will not use the unsecured bond market at current yields; several outlets cited the loan news as one reason for the Nov 13 share-price fall. (Bloomberg via TheStreet and AI Stock Wire, Nov 13)

Platform & Model Tracking

OpenAI (GPT/ChatGPT)

ChatGPT Wallet follow-up — United Airlines bans AI agent bookings (see Top Story 4). Follow-up and supplement (reported Nov 12): OpenAI reportedly closed a $15B round at a $1.25T pre-money valuation, about half its original target — CNBC's David Faber reported on Nov 12, citing sources, that OpenAI had planned to raise up to $30B at a $1.4T valuation, with the first $11B already received on Nov 6 and SoftBank and Abu Dhabi's MGX participating; the company has not confirmed it, the round is not formally closed, and the valuation figure was acknowledged as the pre-money valuation by CFO Sarah Friar in a public interview. The news is dated Nov 12 and was not in the previous issue's list, so it is added here under the CNBC account; the previous issue's unverified item cited only a single-source Analytics Insight "reported" claim. GPT-6.2 remains unreleased and is rumor only — OpenAI's official model catalog had no model ID containing 6.2 as of the evening of Nov 11; a Nov 11 remark by the Codex product lead ("Thursday feels very Codex, maybe a good day for a banked reset") was reposted as a "GPT-6.2 launches Thursday" leak, though the original post named no model; the Nov 12 security keynote also brought no new model. Separate X rumors claim GPT-6.2 Astra/Sol has appeared on Azure and Google Cloud quota pages, with no official confirmation.

Anthropic (Claude)

IPO underwriters reportedly refuse China and Hong Kong investors (see Top Story 2). Background (disclosed Nov 10, first recorded in this issue): Anthropic has determined it has met the Responsible Scaling Policy threshold for "AI substantially accelerating its own R&D" — Amendment No. 2 to the S-1 changed the Nov 2 wording from "may soon reach" to "has determined it has met" the threshold, based on expectations about future models and the overall pace of progress (METR's assessment puts the pace at about 1.9x since October 2025), with no current model individually deemed to have crossed the line; the company says it will assess the impact and publish a further report. Background (support-region policy effective Nov 12, first recorded in this issue): Claude has stopped serving people and entities in mainland China, Hong Kong, Macau, Russia, Iran, North Korea and other listed places, and the restriction extends to companies majority-controlled by people or entities from those places wherever their employees are based; Taiwan is on the supported list. S-1 financial details (summarized Nov 14, from SEC filings): revenue of $33.2B in the first nine months of 2026 (versus $2.5B a year earlier) and an operating loss of $1.3B; purchase and lease commitments of $70.7B falling due within 12 months and $900.9B cumulatively through 2045, almost all for compute and infrastructure; the founders' LLC holds special shares with 50.1% voting power on most matters. No new Claude outage in the window; the Sonnet 4.5 retirement schedule is unchanged (deprecated Sep 30, hard retirement Nov 30, with Sonnet 5.5 as the recommended replacement). Single source: Anthropic published ten recommendations on how to design, run and document safety testing, unverified.

Google DeepMind (Gemini)

Gemini connects to Google Wallet (see Top Story 6). Gemini 4 Hydrogen follow-up benchmarks and distribution — Artificial Analysis Intelligence Index scores 51 at the default high setting (above Argon at 50) and 54 at xhigh, with pricing of $1.50/$7.50 per million tokens and a flat 1M-token context; Google says Terminal-Bench 4.1 is 57% (3.8 Flash was 18.2%) and tasks run 2.5x faster than 3.8 Flash; Antigravity adds an Orchestrator that switches between Carbon and Hydrogen within a session. GitHub Copilot announced on Nov 13 that it will retire six more models on Dec 14, with Gemini 4 Hydrogen going GA in Copilot the same day — from Dec 14, Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3 leave Copilot on all platforms, replaced by Fable 5.2, Haiku 5.5, Gemini 4 Hydrogen, Gemini 4 Carbon and Grok 4.7 respectively, with Kimi K3's replacement still to be announced; annual personal subscribers can keep using Haiku 4.5; Gemini 4 Hydrogen became generally available in Copilot on Nov 13 (whether Fable 5's replacement is 5.2 or 5.1 is disputed across sources, see Unverified Items).

Meta

Single source: privacy researcher Paul Walsh published a technical analysis on Nov 14 alleging Muse can build detailed profiles of people who have never used Muse through accounts connected by others; Meta has not responded in the window (unverified). Single source: according to TestingCatalog, a Muse Chrome extension is now listed on the Chrome Web Store, letting users continue conversations in the browser sidebar and operate web pages, though Meta has not formally announced it (unverified). No new Llama model, pricing or availability release in the window; no new ruling in the five-company lawsuit seeking a freeze on frontier models.

xAI (Grok)

Grok 4.8 still has no official confirmation — Musk said on Nov 10 that pretraining was complete, RL had begun and a soft release to some Grok Bot users was planned later in the week; social posts on Nov 12 claimed a quiet rollout to a small group of Bot users, but there is no xAI press release, model card or API information to confirm it; the 2.5T-parameter and C++-rewritten-stack claims still rest only on Musk's own posts. No Grok 5 release or pricing change in the window.

Microsoft

Microsoft 365 global outage fully restored in the early hours of Nov 14 (follow-up) — the outage that began Nov 12 affected Outlook, Teams, SharePoint, Copilot and Defender XDR; Microsoft says the underlying fix is complete and remaining users' service is restored, with no sign of a cyberattack, and has promised a post-mortem that had not been published as of Nov 14; note that the previous issue said the outage was resolved on Nov 13, while this source puts full restoration in the early hours of Nov 14 (see Unverified Items). Single source: Copilot in Excel's accounting features (reconciliation, drafting journal entries) went generally available worldwide from Nov 13, rolling out in stages through the end of November (Microsoft Message Center MC1536053, via a single source). No new Copilot model or pricing developments in the window.

NVIDIA

Director Stevens' stock sale and the week's market performance (see Top Story 3). The deadline for Huang to accept the House's Dec 1 closed-door testimony is 20:00 ET on Nov 14 (09:00 Beijing time on Nov 15, outside the window), with no report in the window on whether he accepted. Q3 earnings are due Tuesday, Nov 17, hosted by interim CEO Colette Kress, the first earnings report since Huang's departure. No new developments in the CEO search in the window (the FT report on approaching Rene Haas and the board's denial were covered in the previous issue).

Mistral

CEO Mensch briefed the Dublin EU AI summit (see Top Story 8). No new product or funding in the window; the official news page's latest items as of Nov 14 are still the Nov 5 Studio expansion and the Oct 29 Large 4 open weights.

Perplexity

Teammates reached the Mac app on Nov 13, for Max subscribers only ($200/month), following the Nov 5 web launch (the previous issue already reported the Mac app; this item confirms the detail that Pro and lower tiers do not get it). Decider v1.3 and Flash v1.3 were open-sourced on Nov 12 (Apache license, Hugging Face), averaging 91.6% accuracy across 63 internal tasks, with a free window on OpenRouter through Nov 19 (reported in the previous issue; this item confirms the details).

China AI Companies

Alibaba (Qwen)

Qwen4-Max preview has no new release, with early hands-on results and spec details still coming out — the preview (Nov 11 to Nov 25) brought no new version in the window; a Nov 12 roundup of third-party tests reports positive early feedback on building apps from scratch (one coding task finished in 2.5 minutes, versus about 17 minutes for Qwen3.8-Max), but Alibaba has published no official benchmark scores and neither Artificial Analysis nor LMArena had scores as of Nov 12; confirmed specs are a 1M-token context, up to 393,216 output tokens, Token Plan subscription only, no pay-as-you-go, and the model ID goes offline when the preview ends. Claims of four Qwen4 versions and more than 3T parameters are unverified (see Unverified Items).

Baidu (ERNIE)

Baidu World 2026 follow-up — ERNIE 5.5 Flash Preview specs and pricing: 280B total parameters with 14B activated per token and a 1M-token context, live on the Qianfan platform at 1 yuan per million input tokens and 4 yuan per million output tokens, with weights to be open-sourced on Nov 25; the ERNIE Lightning deep-thinking model and ERNIE Wizshot micro-expression model from the same event are also live on the Qianfan API. No new Baidu release on Nov 13–14; Robin Li gave a digital-human interview on Nov 14 responding to the "early start, late arrival" label, a PR statement with no product information.

Tencent (Hunyuan/Yuanbao)

The free access to Hunyuan Hy4/Hy3 ends at 23:59 on Nov 15 — several quota trackers confirm the free period for Hy4/Hy3 inside WorkBuddy and CodeBuddy ends at 23:59 on Nov 15, the same day Zhipu's AutoClaw second-week quota expires. The Hy Realtime voice model was covered on Nov 13, with no new version in the window; Hy4 weights still not released as of Nov 14. A rumored plan for Tencent to cut about 50% of Yuanbao and Hunyuan staff is unverified (see Unverified Items).

Zhipu (GLM) / MiniMax

An NVIDIA developer blog says GLM-5.3 hit 607.8 tokens per second on two DGX Stations — on Nov 13 the blog said two DGX Stations linked by ConnectX-8 RDMA ran the 753B-parameter GLM-5.3 NVFP4 model with vLLM tensor parallelism, peaking at 607.8 tokens per second with 16-way concurrency and 482.90ms time to first token, with the roughly 464GB checkpoint split across two GB300 GPUs (a vendor technical-blog claim). AutoClaw news and quotas: the GLM-5.3-flashX-Legal legal-specialist model is live on AutoClaw (official blog, Nov 13); the second-week quota (50M tokens in China, 100M overseas) expires at 23:59 on Nov 15, and logging in during the third week, Nov 16–22, grants another 100M tokens. Hong Kong stocks: no trading on Saturday Nov 14; Zhipu's intraday fall widened to 11.32% on Nov 13 (MiniMax down 8.31% at the same time), with the final closing drop still unverified (see Unverified Items); MiniMax closed at HK$152.4 on Nov 13, down 8.19% (covered in the previous issue). Single source: JPMorgan cut MiniMax's target price from HK$300 to HK$200 on Nov 12, keeping a Neutral rating (via Longbridge). The claim that Revolut adopted GLM is unverified (see Unverified Items).

ByteDance (Doubao)

Doubao Work opens to third-party models (see Top Story 5). Follow-up on Zhang Yiming's remark that ByteDance cannot catch up with OpenAI and Anthropic until 2029–2030: Bloomberg, citing people familiar with the matter, says Zhang told an all-hands of the Seed team that ByteDance could only hope to compete at the frontier in 2029–2030, with Seed's pretraining dataset orders of magnitude smaller than its top rivals; ByteDance plans a Meta Muse-like personal AI assistant in the coming weeks (the internal "Spell" project is being folded in with the Doubao team) and plans to invest up to $70B in AI data centers. A ByteDance spokesperson did not respond to requests for comment; the content is attributed to sources, not an official statement. No new developments on the reported resignation of AI chip chief Shi Yunfeng (see Unverified Items).

DeepSeek

App API pricing details are out (see Top Story 7). A first consumer paid plan ("Coding Plan") is rumored from frontend code, unverified (see Unverified Items). A former OpenAI engineer's claim that DeepSeek-V4 and five other Chinese models carry NVIDIA Blackwell initialization "fingerprints" is unverified (see Unverified Items). No new model release in the window; the second-round financing figures of about 100B yuan versus 105.5B yuan remain unconfirmed, with no official statement (see Unverified Items).

Moonshot (Kimi)

Kimi 3.1 Pro joins the first batch of third-party models in Doubao Work (see Top Story 5); third-party models were previously enterprise-only. No new release from Kimi itself in the window; reports of Kimi K4 (Q2 2027, 5.4T parameters) and K3.5 are Nov 7 news, with no new developments in the window.

Meituan (LongCat)

No new developments in the window; the LongCat-2.5 weights were open-sourced on Nov 9–10 (already covered).

Manus

No new developments found in the window.

Other China developments

China's AI export-control draft comment period closes Nov 15 — the Ministry of Commerce and Ministry of Science and Technology's Announcement No. 48 proposes adding 40 technologies (23 of them AI), with the control threshold for large models set at training compute of at least 10^26 FLOPs, at least 2T parameters, or a global top-ten ranking, covering weights, checkpoints, source code and architecture; already-open-sourced technology and API/SaaS services are exempt; the comment period runs Oct 16 to Nov 15, and the interim pre-export registration requirement took effect Oct 16. No new policy was issued in the window; this is a deadline reminder.

Policy and regulation

BAAJA follow-up — no new vote in the window, with the Senate in recess until Nov 16; a Nov 14 energy analysis reiterates that the final vote is expected in the week of Nov 16, with the bill letting AI data centers build their own power facilities and limiting challenges by state attorneys general and tribal bodies; no new vote or amendment news. House leaders urged on Nov 13 that four chip export-control bills be folded into the FY2027 NDAA, adding a fourth, the Remote Access Security Act — House committee leaders called on Nov 13 for swift passage of four bills (the AI OVERWATCH Act, the Chip Security Act, the MATCH Act and the Remote Access Security Act), with OpenAI having privately expressed support to congressional offices this week and Congress needing to finish the FY2027 defense authorization bill by Dec 31; the core OpenAI backing of three bills was covered in the previous issue, and the new elements here are the fourth bill and the Nov 13 push (published Nov 13 17:35 EST, about 25 minutes before the window opened, a boundary-time item). Megaspeed follow-up: second-hand reports on Nov 13 added that the Commerce Department paused its Megaspeed investigation at the Justice Department's request and never put it on the Entity List — per the New York Times, the Commerce export-control probe launched in late 2024 was paused and has become a Justice Department criminal investigation, with no new Entity List additions in over a year; but the report itself is dated Nov 9 and the Nov 13 pieces are mostly re-reporting, and the Wyden–Neal letter of Nov 12 covered in the previous issue remains the latest official action, with no new official developments in the window. EU AI Act: no new developments in the window — drafting of the transparency implementation guidelines (Article 50) started Nov 5, with the final version expected in June 2026, and no new official release on Nov 13–14.

Data centers and energy

Nebius, CoreWeave and IREN all fell on Nov 13 (see Capital markets), with Nebius already down about 17% the previous day after cutting its 2026 outlook. Single source: thousands marched in Edinburgh on Nov 14 against AI data centers, demanding a public say before approvals — Friends of the Earth Scotland says 23 projects under review would add 9.1 GW of demand against a current peak demand of 5 GW (a scenario estimate, not a forecast); the innovation minister says the existing planning system can deliver community benefits; only Superpower Daily reported it.

Capital markets

US markets were closed on Saturday Nov 14, with no new trading day; the Nov 13 closes were covered in the previous issue (S&P 500 7,561.67, -0.40%; Nasdaq 25,136.27, -0.91%; Dow 52,347.55, +0.10%). A one-day reversal in chip rotation: AMD closed at $534.06 (-8.75%), Broadcom at $347.01 (-4.76%) and TSMC's ADR at $398.73 (-1.67%) — AMD and Broadcom gave back most of their Nov 12 surge on Huang's resignation, but are still up 18.9% and 18.2% for the week; the Philadelphia Semiconductor Index fell about 3.4–4.2% on Nov 13. CoreWeave seeks at least $4B in investment-grade loans (see Top Story 9). Anthropic IPO (see Top Story 2). OpenAI's $15B round under the CNBC account (see OpenAI tracking). No new large funding disclosures this week beyond the Crunchbase Nov 7–13 weekly list already covered in the previous issue (Mirendil $1B Series A, Mercor $750M Series D); Fractile's $7.2B Series C was announced Nov 11 and covered in the previous issue, so it is not repeated.

Security incidents

FBI ARTEX alert (see Top Story 1). Notion/Fireworks breach update on Nov 13 (follow-up): Notion has notified third-party vendors about leaked API keys, and Fireworks acknowledges that billing data, keys and incomplete model weights were also touched — Notion revoked the leaked Notion API keys on Nov 11 and notified the relevant vendors on Nov 12 about third-party keys found in the chat records; Fireworks says the attackers were a for-profit group aiming to resell LLM tokens (LLMJacking), and its Inference and other zero-data-retention services were unaffected; according to a Nov 14 security briefing, Cursor confirmed the attackers obtained one of its keys sent to Fireworks and used it in an unauthorized Claude Code extension incident in the Cursor marketplace. The Unit 42 autonomous-intrusion report (Cairn with DeepSeek and GLM) was covered in the previous issue, with no new developments in the window; it corroborates the FBI ARTEX alert, see Top Story 1. The Microsoft 365 outage was explicitly not an attack (see Microsoft tracking).

Worth Reading

1."FBI warns that the ARTEX AI hacking tool can reach privileged access in under three hours" — OODA Loop, 2026-11-13

The background and source mapping of FLASH-20261113-001, the source piece behind Top Story 1.

2."Anthropic IPO filing: S-1/A threshold disclosure" — enjoy.dev, 2026-11-13

A line-by-line comparison of the Nov 2 and Nov 10 wordings and the RSP v3.4 definition, the verification source behind the Anthropic tracking item.

3."Checking every NVIDIA Form 4: how much insider selling really happened" — Buteon, 2026-11-13

A tally of all Form 4 filings since mid-2024, concluding Q3 insider selling totaled about $968M, similar in scale to prior years, but 98% concentrated in Stevens with no 10b5-1 plan — the verification source behind Top Story 3.

Unverified Items

1.

Anthropic IPO refusing mainland China and Hong Kong investors: Bloomberg citing people familiar with the matter, via secondary reporting; the prospectus does not state the restriction in so many words, and the outcome of the BVI-trust application is undecided.

2.

Director Stevens' $946M stock sale: the Form 4 data is verifiable, but the reading of "no 10b5-1 plan" and the motive are media inferences; Buteon's Q3 total of about $968M differs slightly from the $946M figure.

3.

DeepSeek's first consumer paid plan ("Coding Plan"): only frontend code clues and community speculation; the company has confirmed no pricing, tiers or launch date.

4.

Tencent planning to cut about 50% of Yuanbao and Hunyuan staff: only a Nov 12 Weibo rumor, with no official response and no cross-checking by credible outlets.

5.

Four Qwen4 versions with more than 3T parameters, an RSI self-recursion mechanism and Huawei Ascend support: claims from self-media, unconfirmed by Alibaba.

6.

Zhipu's final Nov 13 closing drop: only the intraday figure of -11.32% is seen, with the closing price unverified; the previous issue's -10.04% likewise had no source directly stating the close.

7.

Six Chinese models carrying Blackwell initialization "fingerprints": a one-sided research conclusion by former OpenAI engineer Will DePue, whose method has no independent verification and cannot prove where the chips actually are (they may have been used via US cloud providers).

8.

Revolut having about 2,000 engineers use GLM by default, cutting coding time by about 60% and costs by nearly 90%: only a single Chinese self-media repost, with no confirmation from Revolut or Zhipu.

9.

Microsoft 365 outage full-restoration timing: the previous issue said it was resolved on Nov 13, while a dedicated post-mortem puts full restoration in the early hours of Nov 14.

10.

GitHub Copilot retirement replacements: dev.to and Pondero say Fable 5 is replaced by Fable 5.2, while a Medium roundup writes Fable 5.1 — take GitHub's official changelog as authoritative; this piece did not open the official page to check.

11.

Nebius's up-to-$19.4B contract with Microsoft lagging in delivery in New Jersey: only one Nov 14 market analysis, with no other source cross-checking it.

12.

Edinburgh march against data centers: only Superpower Daily reported it; the 9.1 GW of added demand from 23 projects under review is a scenario estimate, not a forecast.

13.

Meta Muse profiling allegation and the Chrome extension: each single source, with no response or formal announcement from Meta.

14.

OpenAI's $15B round: CNBC citing sources, with the company not confirming it and the round not formally closed; the CFO acknowledged the pre-money valuation figure.

15.

Anthropic's RSP threshold determination: based on expectations about future models and the overall pace of progress, not on measurements of current models.

16.

Anthropic's nine-month revenue of $33.2B, operating loss of $1.3B and $70.7B in commitments due within 12 months: relayed from SEC filings via a weekly roundup, not checked line by line against the originals.

17.

The Megaspeed investigation being paused and the company never listed: a Nov 9 New York Times report via secondary reporting, with no new official developments in the window.

18.

BAAJA's two cloture votes (about 18:00 on Nov 16 and about 01:00 on Nov 18): only the Hill Heat newsletter, with no Senate floor schedule to confirm (carried over).

19.

Anthropic IPO pricing Nov 18 and trading Nov 19: the Reuters account; CNBC and the WSJ previously reported Nov 20, the FT says pricing is expected Nov 18, and the official filing leaves the dates blank — take the official announcement as authoritative (carried over).

20.

Anthropic IPO six-times oversubscription: sources cited by Reuters, with Anthropic declining to comment and order numbers still subject to change; final pricing is only confirmed on Nov 18 (carried over).

21.

Grok 4.8 quiet rollout: only social posts and aggregator headlines claiming a limited rollout to Bot users on Nov 12, with no official confirmation from xAI (carried over).

22.

GPT-6.2 launch rumors: only X posts and screenshot rumors; OpenAI's official catalog has no such model (carried over).

23.

DeepSeek second-round financing figures: about 100B yuan (Caijing) versus 105.5B yuan (21st Century Business Herald and others), with pre-money valuation of about 500B yuan versus post-money 615.5B yuan, and no official confirmation (carried over).

24.

Unit 42 autonomous-intrusion report: a single-source chain (the Unit 42 report via CybersecurityNews and Reseller NZ), unverified (carried over).

25.

Zhipu ARR of $1.6B: no new corroboration found in the window, still single source and unverified (carried over).

26.

T-Head's confidential A1 filing: single source (China Fund News via Xueqiu), with the announcement text unverified (carried over).

27.

H3-Image license reading (not applicable in the US, EU, UK and South Korea; companies with annual revenue over $1M need separate written permission): a single-source reading, unverified (carried over).

28.

Epoch AI dataset (15.1 GW of leading data-center compute, 38 GW by end-2028): a single-source dataset, with power figures estimated, unverified (carried over).

29.

Chainstack AI-driven attack: single source, with the incident itself outside the window and the disclosure inside it, unverified (carried over).

30.

CoreWeave Q3 revenue figures: single-source aggregator reporting, with no Reuters or Bloomberg original link (figures fit the earnings-season timing, but unverified) (carried over).

31.

Claude Fable 5.2 fast mode on the Messages API: only one API aggregator briefing, unverified (carried over).

32.

Mirendil $1B Series A, UFORCE $700M Series A, Agile Robots $500M funding, Luminal $25M Series A and Databento $100M Series C: all single source or aggregator relays, unverified (carried over).

33.

Menlo Ventures report data: single source (21st Century Business Herald relaying a China Securities Journal reading of the Menlo Ventures report), with the original report data unverified (carried over).

34.

ByteDance AI chip chief Shi Yunfeng's resignation: single-source reporting, with no official response from ByteDance and the next destination unknown (carried over).

35.

Huang's Nov 18 closed-door briefing: no report of a public reply before the Nov 11 17:00 deadline; he resigned on Nov 12, and whether the briefing will take place or who will attend is unconfirmed; the discussion has now been superseded by the Senate Banking Committee testimony requirement of Dec 3 (carried over, updated).

36.

Whether Lutnick will be removed: Reuters citing sources that the White House is discussing it, with no official response; White House officials said on Nov 12 that Trump has "full confidence" in Lutnick, so the signals conflict and the matter is unconfirmed (carried over).

37.

Huang testifying Dec 3 and Lutnick Dec 10: the Senate Banking Committee summons is relayed by FourWeekMBA from NBC and CNBC, and Lutnick's agreement to appear is a single-source update; the House's Dec 1 closed-door interview required Huang to confirm in writing by 20:00 ET on Nov 13 or face a subpoena, with no report of the outcome in the window; Mast says Huang has agreed to testify before Thanksgiving on Nov 26, with the date still being negotiated (carried over).

38.

Whether NVIDIA approached Rene Haas: the FT report via Reuters, with Reuters saying it could not independently verify it; NVIDIA's board denies having formally contacted any candidate (carried over).

39.

Anthropic IPO total raise figures: about $64.4B in the public offering (Reuters), with totals including the private placement given as "about $90.4B" (Yahoo Finance via Bloomberg) and "over $100B including the greenshoe" (FT) — take the official filing as authoritative (carried over).

40.

The seven co-founders selling about $1.75B of stock: Bloomberg via Pollar, saying each sold 3,378,378 shares for about $250M at the $74 midpoint, about $1.75B in total, with JonesTrading also marketing a block trade of up to $500M; single source, and the piece was published Nov 10 and updated Nov 13, so the timing boundary is unclear (carried over).

41.

Three senators' letters pressing Lutnick: only two secondary sources, with no authoritative reporting confirming the letters' contents (carried over).

42.

Senators Curtis and Blunt Rochester writing to Lutnick about a "remote access" loophole: AI Brief single source, unverified; OpenAI backing the Chip Security Act, MATCH Act and AI OVERWATCH Act: a Nov 12 Bloomberg report via WinBuzzer, a single-source chain (carried over).

43.

Bloomberg's "restricted NVIDIA chips traced to a Shanghai data center": Engadget and Seeking Alpha relaying Bloomberg; the documents do not prove Alice Huang herself took part in chip diversion; Megaspeed has bought more than $4.8B of NVIDIA chips since June 2024, unverified independently (carried over).

44.

LITONG moving 240,000 NVIDIA chips (at least $7.5B) into China: only in a relay of The Information's reporting, with no other independent source, and the claim involves smuggling allegations, so it is marked unverified (carried over).

45.

DeepSeek-linked investors under US investigation: a KuCoin flash report saying companies linked to Alice Huang invested in DeepSeek around Nov 12 and promised NVIDIA chip access, while her Singapore-based Megaspeed International is under US investigation over NVIDIA chip imports worth billions of dollars, with NVIDIA's board also reviewing the matter; the piece itself says the intelligence-link claim "remains only a question". Only a single aggregator flash, with no mainstream corroboration, and the facts and figures are unverified — not suitable as main-text news (carried over).

46.

NVIDIA's Nov 13 closing price figures: Yahoo Finance and Finnhub give $162.85 (-2.01%), while individual quote sources report $161.24, $161.52 or $161.77, likely after-hours or data-vendor differences; the main text uses $162.85 (carried over, updated).

47.

NVDA's Nov 12 closing price figures: ChartExchange recorded $164.27 (-5.45%), versus $166.19 (-4.34%) from Finnhub and several closing roundups — Finnhub is used here (carried over).