NVIDIA Opens an Independent Inquiry as AI Faces Its Power and Permission Limits

Monday, 9 November 2026 | Global AI News Daily Ledger

The week opened with the most sensitive story in AI: NVIDIA's board handed the alleged foreign-intelligence targeting of Jensen Huang to independent directors and outside counsel, days before earnings and a closed-door House briefing. A report put Anthropic behind $3.4 billion of gas-generator orders to guarantee power, Trump told 60 Minutes he doesn't mind AI guardrails, and Apple put its agentic-Siri roadmap on the clock. The builder layer had its own reckoning: John Langford declared Microsoft AI Frontiers over and walked, Alibaba's real-world benchmark showed the best coding agent fully solving 26.4% of a deliberately hard sample of requests, DeepSeek's Harness passed 1.5 million daily users, Qwen Office 2.0 and OpenCode Browse pushed agents into team workflows and signed-in browser tabs, and DARPA started hunting for ways to break digital markets without breaking into computers.

Coverage window Since 8 November, 18:45 IST | Compiled 9 November, 18:45 IST

1. NVIDIA board opens independent investigation into alleged foreign-intelligence targeting of Jensen Huang

NVIDIA's board said 8 November it has initiated an independent investigation - overseen by independent directors with outside counsel - after reporting that founder and CEO Jensen Huang was 'the target of a foreign intelligence effort'. The board said it has no reason to believe national security or NVIDIA secrets were compromised, that Huang fully supports the review, and that it retains confidence in his leadership. CNBC, citing a staff email it obtained, reported Huang acknowledged being unfaithful to his wife and said he is 'proud and grateful to be an American'; NVIDIA told CNBC he remains CEO 'without interruption'. CNBC cited New York Times and Financial Times reporting that US intelligence agencies are concerned a businesswoman affiliated with Megaspeed - a cloud company and NVIDIA customer previously accused of helping restricted Chinese companies acquire NVIDIA chips - was passing information to Beijing. Three House Republican committee chairs have asked Huang to brief lawmakers behind closed doors on 18 November; NVIDIA reports earnings 17 November.

Analysis. The central company of the AI buildout now runs an espionage-adjacent investigation of its own CEO in the same week it reports earnings and faces Congress. Whatever the inquiry finds, export-control enforcement just became personal - and every chip customer list is now a counterintelligence document.

NVIDIA Newsroom: https://nvidianews.nvidia.com/news/statement-from-nvidia-board-of-directors
CNBC: https://www.cnbc.com/2026/11/08/jensen-huang-nvidia-china-spy-congress.html
NBC News: https://www.nbcnews.com/tech/tech-news/nvidia-ceo-jensen-huang-investigation-rcna608148

2. Anthropic reportedly strikes a $3.4 billion gas-generator deal to guarantee data-center power

The Financial Times reported 8 November that Anthropic struck a deal worth about $3.4 billion to secure gas-fired generators capable of producing 1.7 gigawatts - taking over existing heavy-duty generator orders from San Antonio-based OnePWR Solutions in late September rather than acquiring the company, with engineering and construction capabilities included for installation. Reuters could not immediately verify the report, and neither Anthropic nor OnePWR commented. The move fits Anthropic's reported plan to spend more than $500 billion on AI infrastructure over a decade, and lands days after its amended S-1 disclosed roughly $895 billion in compute and infrastructure commitments (covered 3 November) - and in the same week data-center politics dominated the US midterms aftermath.

Analysis. If confirmed, the deal would show a model company securing power equipment as well as chips. Generator orders and installation capability could become part of capacity planning as grid connections constrain expansion.

CNBC: https://www.cnbc.com/2026/11/08/anthropic-generator-deal-ai-data-centers.html
The Next Web: https://thenextweb.com/news/anthropic-3-4bn-gas-generators-onepwr-ft

3. Apple reportedly targets iOS 27.3 for first-party Siri app control, with third-party preview in spring 2027

Bloomberg's Mark Gurman reports (via The Verge, 8 November) that CEO John Ternus is pressuring software chief Craig Federighi to accelerate Siri into an AI agent competitive with OpenAI's Dots and Meta's Muse. Near-term: fine-grained control of Apple's own apps - rotating and cropping in Photos, organizing Safari tabs - expected with iOS 27.3, with some features possibly reaching public builds in December (AppleInsider). Third-party app integration (ordering via Grubhub, booking via OpenTable) is likely a spring 2027 'preview' around iOS 27.5-27.6, ahead of WWDC 2027. Longer horizon: a Siri that tracks goals and pursues them without step-by-step prompting, which Gurman believes could arrive as early as next year. All of this is reporting on plans, not shipped product.

Analysis. Apple is writing its catch-up anxiety into the software calendar - and the sequencing admits the strategy: master its own apps first, open to third parties only when the agent can be trusted with them. A 2027 agentic Siri would arrive two years after the incumbents, into a market where memory and autonomy are already table stakes.

The Verge: https://www.theverge.com/ai-artificial-intelligence/1029552/apple-siri-ai-agentic
MacObserver: https://www.macobserver.com/news/siri-ai-agent-2027-gurman/

4. John Langford declares the end of Microsoft AI Frontiers and leaves

John Langford published 'The End of AI Frontiers' on 6 November, announcing that the ~50-person research group he created inside Microsoft Research - home of the AutoGen orchestration framework and the Fara computer-use agent - has been folded into Microsoft AI and that he has left the company. Langford wrote that the philosophy of Microsoft AI's former-Inflection leadership - 'scaling, and aggressively not research' - was incompatible with AIF's mission, that he 'could not get innovations into use, or even seriously considered', and that every other founding AIF leadership member reached a similar conclusion. He declined to return to MSR, arguing its basic deal - 'do good work and good things happen' - had broken down: 'we did good work and were non-consensually reorged away.' He cited the Full Bandwidth Transformer, Next Latent objective and Dion3 as research whose returns he believes are real.

Analysis. A public autopsy of big-lab research in the scaling era: the team that built AutoGen - the most cited multi-agent framework - is gone because the parent company chose scaling over research. Where Langford and his colleagues land will tell us where agent research goes next.

hunch.net: https://hunch.net/?p=13764006
The Brief: https://www.thebrief.news/en/standard/article/41133/microsofts-ai-frontiers-research-group-closes-after-leader-departs

5. Trump tells 60 Minutes he doesn't mind AI guardrails, signals openness to working with Democrats

President Trump said in a 60 Minutes interview aired 8 November that he doesn't mind AI guardrails and suggested he could work with Democrats on safeguards - while warning that regulation must not leave the US behind China (Reuters). The comments land as Democrats, fresh from their midterm sweep, prepare hearings and a surge of AI bills, and as the administration's own DOJ stands up its Super Intelligence Task Force (covered 6 November). Trump framed AI risks as manageable, a notable tonal shift given his administration's deregulatory posture and the election's data-center backlash.

Analysis. A bipartisan window for federal AI legislation is opening for the first time - with the White House signaling conditional openness and the new congressional majority eager to move. The China-competition caveat is where the negotiations will live or die.

Reuters: https://www.investing.com/news/world-news/trump-says-ai-risks-can-be-managed-signals-openness-to-guardrails-5010374
Startup Fortune: https://startupfortune.com/trump-tells-60-minutes-he-doesnt-mind-ai-guardrails-as-democrats-prepare-hearings/

6. Alibaba's deliberately hard coding benchmark: the best tested endpoint fully completes 26.4% of tasks

Alibaba researchers built a benchmark from real-world development requests rather than synthetic puzzles, and the results cut against leaderboard optimism: the top coding agent met about 70% of stated requirements but fully completed only 26.4% of tasks perfectly - even as comparable models score above 99% on some older benchmarks (The Decoder, 8 November). The gap comes from the messy parts synthetic tests strip out: ambiguous requirements, multi-file context, environment quirks and integration work that real tickets carry. The finding echoes METR-style observations that benchmark saturation has outpaced deployment reality. Researchers favored requests that had previously required repeated clarification or rollbacks. The test fixes one request per task, excludes continuing user correction, and relies on AI judges that can err. This is not an estimated failure rate for all developer work.

Analysis. A useful counterweight to app-building demonstrations: meeting most requirements is different from completing a task. The selected sample and grading limits matter, so buyers should test their own workflows rather than generalize the headline percentage.

The Decoder: https://the-decoder.com/ai-coding-agents-still-need-human-cleanup-on-nearly-three-out-of-four-tasks-in-a-real-world-test/
DEV Community: https://dev.to/piekwerk/alibaba-measured-it-the-best-coding-agent-meets-70-of-requirements-and-finishes-26-of-tasks-52be

7. DeepSeek's Harness agent platform passes 1.5 million daily users, burning ~92 trillion tokens a day

DeepSeek disclosed that its open-source agent platform Harness hit 1.54 million daily active users on 4 November - less than three months after launch - processing roughly 92 trillion tokens daily, about 60 million per user (Bloomberg, via Techmeme; Business Times). It is the first usage disclosure for the platform and positions DeepSeek beyond model releases into the agent-runtime layer, where Moonshot's Kimi Hosted Agents (6 November edition) and the US labs are also racing. SCMP frames the open question as whether DeepSeek can challenge OpenAI and Anthropic in agents as Harness adoption compounds.

Analysis. If sustained, the company-disclosed usage would make Harness a significant agent-runtime platform. Token consumption alone does not establish task quality, revenue or a fair comparison with paid agent services.

Techmeme: https://www.techmeme.com/261108/p12
Business Times: https://www.businesstimes.com.sg/companies-markets/deepseeks-ai-agent-platform-hits-1-54-million-daily-users

8. Alibaba launches Qwen Office 2.0 with multi-person agent collaboration

Alibaba released Qwen Office 2.0 on 9 November with what it calls industry-first multi-person agent collaboration: teams of agents that coordinate on office workflows, automated tasks that keep running in the cloud after the user's PC shuts down, and human-AI co-creation features (Aibase, ReadmeX). One showcased pattern: a message from a manager can trigger agents to start work automatically. The launch extends Alibaba's agent push from models (Qwen family) into the collaboration layer, a day after the company's real-world benchmark showed how far coding agents still have to go - both stories arriving from the same research organization within 48 hours.

Analysis. Cloud-persistent, manager-triggered agent teams bring coordination and access control into the office-product layer. The launch describes an ambitious workflow, but does not by itself establish comparative reliability against competing systems.

Aibase: https://www.aibase.com/news/32184
ReadmeX: https://readmex.ai/story/2-0-agent-team-agent-86e4847a

9. OpenCode Browse hands coding agents your logged-in browser

OpenCode launched Browse (product page live, coverage 7-8 November): a Chrome extension and companion tooling that lets AI coding agents operate inside your real, signed-in browser tabs - reading pages, clicking through flows, testing web apps as an authenticated user. The design keeps humans on sensitive actions: logging in and payments stay manual while the agent drives the rest. For web developers it closes the loop between writing code and verifying it in a live session, a gap that has kept browser-based testing of authenticated flows one of the most manual parts of the agent workflow.

Analysis. The browser is becoming the agent's second hands, and 'agent drives my logged-in session' is a genuine capability unlock for web work - with an obvious trust boundary that will be tested the first time an agent's tab access is abused. The Postman hook lesson (8 November edition) applies directly.

OpenCode: https://opencode.ai/product/browse
Chrome Web Store: https://chromewebstore.google.com/detail/browse/mfnicocicmmlkpjnaffgihfjhdgjkdjg

10. DARPA launches program to find ways attackers could break digital markets, not just computers

DARPA opened its Intrinsic Market Resilience (IMR) program (solicitation PS-26-138, reported 8 November by Business Insider): a research effort to find weaknesses in the rules of digital markets - electronic trading, ad exchanges and other online systems - that attackers could exploit to disrupt economic activity without breaching any computer. The premise: market rules themselves are an attack surface, and AI agents make probing them at scale cheap. The program invites researchers to model and stress market mechanisms before adversaries do.

Analysis. As agents get cheaper and more capable, 'hacking the rules' replaces 'hacking the machines' as the systemic risk - DARPA is formalizing what the agent-security community has been warning about since automated systems started trading, bidding and negotiating with each other.

DARPA: https://www.darpa.mil/research/programs/intrinsic-market-resilience-imr
Business Insider: https://www.businessinsider.com/darpa-wants-to-find-weak-points-digital-markets-2026-11

Scope and sourcing notes

Ten distinct developments, none covered in prior editions. Story 1's board statement is NVIDIA's own; the affair admission comes from a staff email CNBC obtained, and the Megaspeed and Beijing-passage allegations are attributed to New York Times and Financial Times reporting - none are adjudicated, and the board says it has no reason to believe secrets were compromised. Story 2's $3.4 billion is FT-reported and unconfirmed by both companies, preserved as such. Story 3 describes reported plans, not shipped features. Story 5's DeepSeek figures are company-disclosed. Tencent's Hy4 (launched 4 November) was weighed and dropped as stale relative to fresher items. The Suno suits (covered 8 November), METR's Anthropic assessment (covered 6 November), the OpenAI-Hacktron partnership (covered 8 November) and NVIDIA's Vera Rubin memory claims (covered 8 November) reappear in today's digests and are not repeated. Also dropped: Microsoft's live speech-editing patent filing, the AI-consciousness essay, the Inner Mongolia data-center visit, and CNBC's revisit of agent-teamwork research. No public X post met the evidence threshold, so no Firsthand signal item is included. Prior editions remain unchanged. Additional benchmark context: Co-Design Bench favors historically difficult requests, uses fixed requests and AI judges, and is not a representative failure-rate estimate for all developer work.

Bring agentic AI to your classroom

TechSambad's AI Learning Pilot puts agentic AI in front of your students for one module, not a semester: a live 90-minute taught session inside your timetable, an AI-built pre-class reading brief mapped to your module, a question bank with grading rubrics, a student takeaway pack, and a written outcome report. Run it once and judge it on what your students can do afterwards - Rs 2,000 + GST for the college, no obligation.

https://www.techsambad.com/p/ai-learning-pilot.html?m=1