Morning Brief: Tuesday, November 10

Seventy-four feeds reporting. Two weeks. 5,779 items reduced to what follows. (what we track, how we crawl, subscribe)

Tuesday is about the off switch and who gets to hold it. OpenAI published Control Room, an internal system for tracking and stopping model workloads across its own research infrastructure, on the same day the Commerce Department's chip inquiry into Nvidia was reported halted after the company intervened. One control plane is technical, the other political; both are answers to the question of who can stop a running machine.

The Nvidia story crossed from personal scandal into export policy this cycle. A short seller had flagged the CEO's ties to a suspected Chinese spy six months before it broke; the Commerce inquiry was halted after Nvidia's intervention; the President backed Huang publicly while the plaintiffs' bar organized. Read the hardware announcements in the same window as the market pricing that in – Google unveiled two more specialized AI chips, and Huawei claimed a fully domestic PC. When one vendor's internal governance becomes a supply risk, everyone else's chip roadmap is a hedge.

Control Room is a frontier lab applying to its own infrastructure what CrowdStrike started selling to enterprises a day earlier, and the arXiv batch underneath it is unusually dense on the same seam from the outside: external model-agnostic oversight, two benchmarks on security invariants and shadow invocations in MCP-style runtimes, least-privilege auditing of tool calls through intent graphs, and a governance paper on how you would even constitute an independent verifier. The products assume you already know which run to stop. The papers are still arguing about how you would tell.

Top (5-7 min)

Control Room: A central system to track and stop model workloads across our research infrastructure: OpenAI, 2026-11-09. A lab building a kill switch for itself, not for its customers. Start here; the rest of the brief is this idea at other layers.
Commerce Dept. chip inquiry was halted after Nvidia intervention: Slashdot, 2026-11-09. The export-control process as the actual control plane over compute, and a vendor reaching into it.
A short seller flagged Nvidia CEO's ties to a suspected Chinese spy six months before the scandal: TechCrunch, 2026-11-09. The signal existed and the market ignored it for two quarters. Worth reading as a failure of diligence rather than of intelligence.
Enmo Safety: External, Non-invasive, Model-agnostic Oversight for Off-the-Shelf Language Models: arXiv cs.AI, 2026-11-10. Oversight that does not require access to the model. The structural counterpart to Control Room, which does.
Git considers voting to determine its LLM policy: LWN, 2026-11-09. A project with no governance body considering one in order to decide what machine-written contributions it will take.
Anthropic unveils Claude Fable 5.2, Mythos 5.2 with coding gains and lower prices: InfoWorld, 2026-11-10. Capability up, price down, in the same release. The cost curve is the story more than the benchmark.
Anthropic joins AI's natural gas frenzy, buys $3.4B worth of turbines: TechCrunch, 2026-11-09. What a price cut costs upstream. A model vendor buying generation equipment is a vertical-integration move, not a procurement one.

Themes

The off switch gets an owner

Control Room, CrowdStrike's commercial kill switch, and Grafana's always-on monitoring agents are three implementations of stopping things, built for three different principals: the lab, the enterprise security team, and the operator. The arXiv cluster underneath is working the measurement problem none of the three solve – what condition should trigger the stop, and who is allowed to see enough to evaluate it.

Can CrowdStrike's new 'AI kill switch' shut down rogue AI agents?: Slashdot, 2026-11-09. Containment as an endpoint product feature. The question mark is doing real work.
From Tool Connection to Execution Control: Benchmarking Security Invariants in MCP-Style Agent Runtimes: arXiv cs.AI, 2026-11-10. Invariants a tool runtime should hold, stated as a benchmark instead of a blog post.
ShadowBench-MCP: Measuring Shadow Invocations in MCP-Style Tool Routing: arXiv cs.AI, 2026-11-10. Tool calls the operator did not know happened. You cannot stop what you are not seeing.
Least-Privilege Auditing of Agent Tool Calls with Intent Graphs and Reasoning-Guided Escalation: arXiv cs.AI, 2026-11-10. Least privilege applied per tool call rather than per agent, with an escalation path.
LiveAsset: Learning Intent Assets from Corrective Interactions for Agent Runtime Control: arXiv cs.AI, 2026-11-10. Runtime control learned from the corrections operators were already making by hand.
What Do Collusion Probes Read? Three Pitfalls in Evaluating Multi-Agent Collusion Monitors: arXiv cs.AI, 2026-11-10. A critique of the monitors, not the agents. The most useful paper here if you are buying one.
ErosionBench: Probing Alignment Decay in Enterprise Multi-Agent Systems: arXiv cs.AI, 2026-11-10. Decay over a deployment's lifetime rather than failure at launch. Matches what operators actually report.
Key Institutional Considerations for Designing Effective Independent Verification Organizations: arXiv cs.AI, 2026-11-10. Who audits, under what mandate. The institutional layer the product announcements skip.
Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety Auditing: arXiv cs.AI, 2026-11-10. Auditing when neither the model nor the eval can be disclosed. Read beside the paper above.
Run cost-effective, always-on agents to monitor your environment: the engineering behind Watchers: Grafana Labs, 2026-11-09. The engineering writeup, including the cost model. Useful precisely because it is not a safety pitch.
Where is all the AI safety research uplift?: Alignment Forum, 2026-11-09. Asks whether capable models have actually accelerated safety work. An uncomfortable question posed from inside the field.
AI Security: Schneier on Security, 2026-11-09. Short framing piece. Separates securing AI systems from using AI for security, which most vendor copy conflates.

Nvidia as a policy risk, not a supplier risk

Everything in this cluster is downstream of one person's conduct reaching the export-control apparatus. The chip announcements are the part worth tracking after the news cycle ends: a concentration risk that was always structural is now visible enough that buyers are funding alternatives.

Trump backs Nvidia CEO Jensen Huang amid Chinese spy affair investigation: Slashdot, 2026-11-09. Political cover extended while an investigation is open.
Jensen Huang's affair with alleged Chinese spy puts Nvidia's China ambitions under scrutiny: TechCrunch, 2026-11-09. The China revenue line as the thing actually at stake.
The race is on to sue Nvidia: TechCrunch, 2026-11-09. Securities litigation forming. The timeline in the complaints will be more precise than the reporting.
Is Jensen Huang's job on the line? Yale's Jeffrey Sonnenfeld estimates it's "about a 50-50": TechCrunch, 2026-11-09. A governance scholar putting a number on it. Treat the number as commentary, the reasoning as useful.
Google unveils two new specialized AI chips to widen its challenge to Nvidia: Slashdot, 2026-11-09. Second-source pressure, announced into the gap.
Google claims its new AI software cuts inference latency by over 90% — but there's a catch: The New Stack, 2026-11-09. The catch is the lock-in. Read with the chip announcement above.
Huawei says it's made a 100% Chinese PC: Slashdot, 2026-11-09. Decoupling claimed at the whole-system level. Verification pending, as always with these.
What Nvidia CEO's affair with a suspected Chinese spy means for the AI race: WBUR, 2026-11-09. Broadcast segment. Useful as a read on how the story lands outside the trade press.

Provenance moves to the governance layer

Last week Asahi Linux relicensed in response to LLM-assisted license laundering – a maintainer acting unilaterally at the license layer. This week Git is considering a vote, which is the same problem escalating to a question about who decides. The arXiv side is working the measurement end, and it keeps returning the same answer: detection is calibration-limited, so policy is going to have to carry the weight.

Hide your slop PR descriptions behind <details><summary>: Simon Willison, 2026-11-09. A convention, not a policy. The small end of the same maintainer problem Git is voting on.
Audit Before You Merge: Provenance, Probing, and Continual LoRA Composition: arXiv cs.AI, 2026-11-10. Provenance at adapter-merge time. The weights analogue of a pull request review.
How Short Is Too Short? Length-Conditional Calibration and Abstention for Detecting AI-Generated Text: arXiv cs.AI, 2026-11-10. Detectors that decline to answer below a length threshold. The honest version of the detection pitch.
OpenMarkX: An Extensible Platform for Watermarking Open-Source Model and Agent Outputs: arXiv cs.AI, 2026-11-10. Watermarking for models anyone can modify, which is the hard case.
What Does a Copyright Budget Certify? An Exact Threshold for Per-Token Meters, and a Length-Independent Certificate for Selection: arXiv cs.AI, 2026-11-10. What a per-token copyright accounting scheme actually guarantees. Mostly a negative result, usefully stated.
Soft Contamination Means Benchmarks Test Shallow Generalization: arXiv cs.AI, 2026-11-10. Contamination short of memorization still inflates scores. Relevant to every number in the model-release coverage below.
State Media at Retrieval Time: A Large-Scale Audit of Russian State-Aligned Sources in AI Chatbot Responses: arXiv cs.AI, 2026-11-10. Provenance failure at retrieval rather than training. A different injection point for the same problem.

Scan

Energy, data centers, and the buildout

The climate case for AI has not been made: MIT Technology Review, 2026-11-09. The strongest piece in this section. Treats "AI will solve climate" as a claim requiring evidence and finds none presented.
The backlash about data centers might really just be about electricity: MIT Technology Review, 2026-11-09.
Fervo now has $12B in orders for its geothermal power tech: TechCrunch, 2026-11-09.
Panic over cloud capacity is out of hand: InfoWorld, 2026-11-10.
How much energy does building distro packages use: Lobsters, 2026-11-09. Small-scale measurement, carefully done. A useful corrective to the everything-is-datacenters framing above.

Policy and society

Nancy Pelosi's replacement is ready to regulate AI: MIT Technology Review, 2026-11-09.
We're asking an appeals court to require a warrant for ALPR searches: EFF Deeplinks, 2026-11-09.
What the Flock? Stories to understand the national controversy around the AI cameras: WBUR, 2026-11-09.
Mass. gave towns money to cope with casinos. Some spent it on Flock cameras: WBUR, 2026-11-09. Local reporting that does the thing the EFF filing argues about: follows the money into specific municipal purchases.
Why fewer people are challenging bad patents: EFF Deeplinks, 2026-11-09.
U.S. Supreme Court deals final blow to corporate human rights accountability: EFF Deeplinks, 2026-11-09.
Pluralistic: America cannot undo Trumpism: Pluralistic, 2026-11-09.
Closing the AI guidance gap in education: OpenAI, 2026-11-10.
Misinformation Speaker Series: Reclaiming the Internet: Northeastern events, 2026-11-10.

Research and math

Improving Analytic Constants via Interpretable AI-Guided Lean Proof Refinement: arXiv cs.AI, 2026-11-10. Machine-guided refinement of a constant, with the proof checked in Lean. The verification story this brief keeps asking for.
ProofDiff: Auditable Argument-Level Comparison of AI-Generated Mathematical Proofs: arXiv cs.AI, 2026-11-10.
BiVerify: Closing the Loop Between Backward and Forward Verification for Mathematical Self-Correction: arXiv cs.AI, 2026-11-10.
An Improved Algorithm for Quantified Boolean Formulas: arXiv cs.AI, 2026-11-10.
Our Survey of Early Career Mathematicians: Terence Tao, 2026-11-09.
Multimodal cell foundation models require synergistic integration beyond alignment: Nature Machine Intelligence, 2026-11-09.
Did Life Begin in an RNA World?: Quanta Magazine, 2026-11-09.
Predictably Unpredictable Inspections: MIT calendar, 2026-11-09.
LIDS Seminar: Tim Roughgarden: MIT calendar, 2026-11-09.
BC-MIT Number Theory Seminar: MIT calendar, 2026-11-10.
Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context: Hugging Face Blog, 2026-11-09.

Hardware and radio

Building a networking switch ASIC chip: Lobsters, 2026-11-10.
Reverse engineering a vintage diode-transistor logic chip: Lobsters, 2026-11-09.
Reverse engineering a 1960s diode-transistor logic IC: Hackaday, 2026-11-09.
Hacking the Sidekick in 2026: Hackaday, 2026-11-10.
From phone, to SDR, no extra hardware: Hackaday, 2026-11-09.
SDR-Receiver: A native macOS SDR app with 25+ decoders, now including Loran-C / eLoran: RTL-SDR, 2026-11-10.
Decoding TETRA voice on a Retevis C62 handheld with custom OpenRTX firmware: RTL-SDR, 2026-11-10.
ESP-WiFi-Jam-Detector: A standalone 2.4 and 5 GHz interference detector on an ESP32-C5 board: RTL-SDR, 2026-11-10.
Boston Dynamics adds wheels to Atlas humanoid: Slashdot, 2026-11-10.
Epic Games comments on planned Linux distribution support: Slashdot, 2026-11-09.

The crawler zoo

Twelve new user agents logged since yesterday, and the week's report puts the total at 53,937 requests. The MCP-survey and agent-readiness crawlers are the same census impulse as the MCP benchmarks above, run from the other side of the wire.

Zoo Report, Week 45, 2026: 53,937 bot requests: The Crawler Zoo, 2026-11-09.
New arrival: NoClick-MCP-Survey: The Crawler Zoo, 2026-11-10.
New arrival: AgentReadinessScanner: The Crawler Zoo, 2026-11-09.
New arrival: bot-research-checker: The Crawler Zoo, 2026-11-10.
New arrival: CloudShuttle-Scout-HTML: The Crawler Zoo, 2026-11-10.
New arrival: MachineWitnessBot: The Crawler Zoo, 2026-11-09.
New arrival: agent-orchestrator-icon-fetch: The Crawler Zoo, 2026-11-09.

Boston

Boston officials announce new housing proposals to entice developers: WBUR, 2026-11-09.
Beacon Hill Democrats had a great election night, too: WBUR, 2026-11-09.
Abnormally dry conditions, warm temps intensified wildfire in N.H.'s White Mountains: WBUR, 2026-11-09.
Plymouth woman IDs herself in 1974 image from famed Boston photographer: WBUR, 2026-11-09.
A new historical fiction novel returns to the Cocoanut Grove nightclub fire: WBUR, 2026-11-10.

Tail

Boeing plans to cut 5% of commercial workforce: Leeham News, 2026-11-09.
Embraer's India ambitions face test of airline economics: Leeham News, 2026-11-09.
Review: elements of a Forth system (Gforth Manual): Pinboard, 2026-11-09.
The Forth Methodology of Charles Moore, by Jeff Fox: Pinboard, 2026-11-09.

Feed silences and verification

One hundred eighteen feeds were polled this morning: 61 returned new content, 56 answered 304 Not Modified, and one errored. Seventy-four published something inside the fourteen-day window.

www.bornholt.dev is the error. It has now failed to resolve on 168 consecutive polls since 2026-05-26 and has never once succeeded from this host, so the feed is removed from the publishing count rather than reported as a gap. The entry stays in the registry pending a decision on whether the domain moved or went away.

Hacker News has answered 304 on every poll for fifteen days and last contributed an item dated 2026-10-26. That is a bug on our side: the cached validator for the Firebase endpoint is stale and we keep sending it, so the source is pinned rather than quiet. No Hacker News items appear in this brief, and none should be read as absence of activity there.

The warehouse holds 69,523 items, of which 3,651 carry no date and are therefore invisible to every window query in this brief, including the item count above. That is unchanged from last week and remains the largest known gap in these numbers.

Every URL above came from data/warehouse.db. Nothing was constructed from a title. Slashdot links had their utm_* feed parameters stripped; no other URL was edited.

The brief carries 157 external links plus one self-link to our own feed. All 158 were verified through crawler/fetch-gated, which runs the blocklist, then robots.txt, then the throttle, and issues no request when the gate denies. Nothing was denied: all 158 were allowed and fetched. Verification took nine minutes parallelized by host, and arXiv accounts for most of that – 37 of the links are arXiv abstracts, serialized on one host by its Crawl-delay: 15.

Results: 151 returned 200. Seven returned 403 to our user agent, and all seven are bot-challenge behavior rather than dead pages: openai.com (five links, including the Control Room post at the top of this brief), lwn.net (the Git article is subscriber-only, marked [$] in LWN's own feed), and netflixtechblog.com (Medium). All seven also returned 403 to a GET, so the refusal is identity-based, not method-based. No link returned 404 and no dead link was found.

Build provenance

build: 2026-11-10 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.4) | feeds: 118 polled today (61 ok, 56 not-modified, 1 error), 74 publishing in window | items-considered: 5779 (14d, incl. 3536 arXiv cs.AI) | warehouse: 69523 items (3651 dateless, see feed silences) | published: 158 URLs (158 allowed by gate, 0 denied; 151 HTTP 200, 7 bot-challenge 403 disclosed; 0 HTTP 404, 0 dead links found) | atom: 157 entries, RFC 4287 valid