- On Friday, November 13, the FAA halted every US departure for four hours and 21 minutes, three days after its AI traffic-flow tool went live at all 22 en route centers. More than 10,000 flights were delayed and at least 4,000 canceled.
- On Thursday, November 12, Nvidia said Jensen Huang had resigned as chief executive, effective immediately, and gave no reason. CFO Colette Kress took over on an interim basis while roughly $1.7 trillion came off the stock's market value across six losing weeks.
- On Monday, November 9, Fireworks AI widened a breach it first disclosed on October 10 to include billing records, customer datasets, and partial model weights, along with Notion AI chat transcripts that sat outside zero data retention.
- On Tuesday, November 10, researchers published findings that 35,578 apps built with Grok opened to anyone holding the link, with no sign-in, among them an ICU patient's lab results and a hospital roster of more than 3,500 staff.
- On Thursday, November 12, Edge0 published open weights for a 145-billion-parameter model that runs on a single 16 GiB GPU, and Google shipped Gemini 4 Hydrogen at double the introductory price of the model it replaces.
- For the frame underneath all of this, start with where generative AI turns into functional AI.
This is the November 16, 2026 edition, covering the week of November 9 to 15. The through-line is not what the machines did this week. It is that almost nobody could produce a record of what they did. An airspace regulator, a chip company's board, an inference provider, a model vendor, and a volunteer maintainer all faced the same question in different clothes, and all of them answered with an assertion rather than a log.
That is a market fact rather than a philosophical one. Logging, tracing, and provenance are cost centers: they slow a launch, grow a bill, and produce evidence that can be used against you later, so the incentive to build them properly points the wrong way right up until the week you need them. This was a cluster of those weeks. Notice who paid to fix it: a nonprofit foundation in Salt Lake City, two overworked maintainers, and security researchers nobody hired. The companies with the balance sheets issued denials.
The Big Story: the FAA stopped 10,000 flights, and three officials gave three answers
This story is not about whether an AI tool grounded American aviation. It is about the fact that four days later nobody can prove either way.
The Federal Aviation Administration issued a nationwide ground stop on departing traffic at 3:54 p.m. Eastern on Friday, November 13, 2026, and lifted it at 8:15 p.m., four hours and 21 minutes later. FlightAware counted more than 10,000 delays and at least 4,000 cancellations, the Associated Press put the number of affected passengers around 1.5 million, and Flightradar24 recorded 4,042 flights already airborne toward US airports when the order came down. Al Jazeera's running account of the day is here, and CBS News covered the lift and the White House response here.
Three days earlier, on November 10, the FAA had finished rolling out SMART, an AI traffic-flow system built by the Boston firm Air Space Intelligence under a contract worth about $875 million over 12 years, to all 22 of the country's air route traffic control centers. SMART ingests roughly 200 data streams, including controller staffing, predicts congestion, and recommends reroutes that specialists forward to the command center in Warrenton, Virginia. If that sounds like an agent, it is: software that takes in state, decides something, and hands the decision to a person who is now the slowest part of the loop. Coverage of the rollout timing is here.
Then the answers diverged. Transportation Secretary Sean Duffy said switches and routers had been overloaded with data, producing inconsistent readings, and that the system was shut down deliberately; radar and communications, he said, were never fully lost. Air Space Intelligence said the ground stop was not based on or caused by SMART, any of its products, or its integration with the FAA, a denial documented here. Homeland Security Secretary Markwayne Mullin, asked the same question, said you cannot rule anything out, as reported here. Investigators are reportedly examining whether a software update contributed. The FBI does not believe a cyberattack was involved.
What has not appeared, from any of the three, is the thing that would settle it: a timestamped record of what SMART recommended in the hours before 3:54 p.m. A $875 million system deployed to 22 facilities should answer that in minutes. The FAA's own performance numbers for SMART, 180 flights rerouted since the September pilot and about 10 minutes and 80 gallons of fuel saved per flight, are agency-reported and unaudited.
"We cannot rule it out" and "it was definitely not us" are both answers you give when you do not have the tape.
What it means: The procurement market for agentic systems in safety-critical infrastructure just got its first real test case, and the lesson buyers will take is about evidence rather than capability. More than two dozen lawmakers had already asked the FAA on October 9 to pause the earlier rollout; they now have a four-hour, 1.5-million-passenger exhibit and no decision log to examine. Expect the next round of these contracts to price auditability as a line item. For vendors the lesson is uglier: a denial you cannot substantiate invites the subpoena that a published trace would have pre-empted.
What's coming: Watch for two things. First, whether any root-cause report is published at all, and whether it covers SMART's recommendations or only the telecom layer, because the scope of that report is the political decision disguised as a technical one. Second, watch Senator Tammy Duckworth, who has already demanded a thorough investigation. If Congress writes a logging requirement into FAA reauthorization language, every agentic infrastructure vendor in the country inherits it by 2028.
Nvidia's CEO resigned with no reason given, and $1.7 trillion repriced around the silence
The most load-bearing company in the AI buildout changed leaders in a press release that explained nothing.
Nvidia announced on November 12, 2026 that Jensen Huang had resigned as president, chief executive, and board member, effective immediately. Colette Kress, the CFO, becomes interim CEO while keeping the finance job, and Harvey C. Jones was named the company's first-ever board chair. The release, which you can read here, thanks Huang for immeasurable contributions and gives no reason for the departure. Reporting over the following day tied the exit to personal-conduct allegations, covered by CNBC here.
The market did the disclosure the company did not. Nvidia closed the week down roughly 13%, its worst since April 2025 and a sixth straight weekly loss, taking about $1.7 trillion off its market value and pushing it below a $4 trillion capitalization for the first time since July 2025. CNBC's summary is here.
What it means: Every compute commitment signed in the last two years was underwritten partly by confidence in one person's roadmap judgment, and that collateral was just marked down without a replacement value being published. An interim CEO who is also the sitting CFO is a holding pattern, not a plan, and holding patterns are expensive when your customers are negotiating multi-year capacity deals. The follow-the-money read is simple: anyone with an unsigned Nvidia commitment got leverage this week, and anyone holding a signed one got a worse counterparty than the one they diligenced.
Fireworks AI's breach kept growing, because the first answer was "no evidence"
This is what happens when an incident's scope is bounded by the logging that existed at the time rather than by what the attacker did.
Fireworks AI, which hosts open-weight models for other companies to run inference on, detected unauthorized use of internal credentials at 9 a.m. Pacific on October 6, 2026 and published an initial bulletin on October 10 describing exposure of environment variables in evaluation jobs at a limited number of accounts. On October 25 it said it had found no further evidence of unauthorized access. On Monday, November 9, it disclosed that an attacker had reached its development machines through a third-party cloud provider on October 23 and 24 and run a workload that exposed more: resource metadata, billing addresses and invoice history, and a subset of incomplete model weights and customer datasets. The Hacker News has the consolidated timeline here and BleepingComputer covered the expansion here.
The sharpest detail belongs to a customer. Notion, which served open-weight models through Fireworks, first believed only synthetic test data was involved, then found real Notion AI chat transcripts used for its Insights feature. Those transcripts were mostly unattributable to specific users or workspaces, which is another way of saying Notion could not determine whose conversations they were. It disabled Insights, stopped serving models through Fireworks, and told affected customers to rotate anything sensitive pasted in during the attack window. Fewer than 1% of workspaces were affected. TechCrunch's account is here.
What it means: "Zero data retention" is a contractual promise about a pipeline, and this week it turned out there was a side channel feeding an analytics feature. If you buy inference through a provider who buys compute from someone else, your data inherits the weakest logging in that chain, and you will learn its shape only during an incident. Fireworks attributes the activity to a financially motivated group that resells stolen model access, which tells you the economics: model inference is now a commodity worth stealing, and the cheapest way to steal it is through the plumbing rather than the model.
The theme of the week was nobody having a record. Your own automations can have one from day one. Tell BYOBot what you want to automate and get a step-by-step spec back, logging included.
35,578 Grok-built apps were readable by anyone, and the dispute is over a preselected checkbox
Two researchers with no commercial relationship to the vendor produced the most complete audit of an AI app builder that anyone has published.
On November 10, 2026, Asaf Aprozper of Claroc Research published findings on apps built with Grok, the assistant from SpaceXSI, formerly xAI. The team probed 43,551 published app addresses, found 37,571 still online, and found that 35,578 of those, 94.7%, opened with no sign-in at all. A sample suggested 650 to 800 held private data. The contents are the story: an ICU patient's lab results, a hospital staff roster of more than 3,500 employees, a trucking company's driver list with personal emails, a timesheet app storing 18 accounts' passwords in plain text, and 24 apps built by SpaceXSI and X staff including a confidential onboarding plan, a sales territory map covering 594 accounts, and an advertising proposal marked confidential and addressed to a US federal agency. The full writeup is here, with secondary coverage here and here.
SpaceXSI's response was that Grok apps are private until a creator shares them, that publishing takes several steps, and that employee accounts default to "Only you." The researchers replied that in the account they used, the share panel offered "Anyone with the link" with that option already selected. The company took down its own staff's 24 apps. It did not change the default for everyone else, and the researchers began notifying affected organizations themselves.
What it means: A default is a security control, and the only people who can verify one from outside are strangers scanning your URL space on their own time. That is the shape of the week: the most rigorous audit of a commercial AI product came from unpaid researchers, and the vendor's rebuttal described intended behavior rather than shipped behavior. If your team builds internal tools with a hosted app builder, the action item is immediate: open one published app in a logged-out private window and see what loads.
Two volunteers hold up a load-bearing project, and a quarter of incoming pull requests are machine slop
The people absorbing the cost of unverifiable AI output this week were not companies. They were maintainers.
At KubeCon + CloudNativeCon North America in Salt Lake City, which ran November 10 to 13, a keynote on Ingress NGINX made the arithmetic plain: two maintainers, a project running inside an enormous number of Kubernetes clusters, and an inbox filling faster than two people can triage because the submissions are now generated. Google's open source roundup for the week, with links to the talk and the surrounding research, is here.
The numbers behind the complaint come from LF Research, written by data scientist Sam Boysel with Anni Lai and produced with Red Hat, Futurewei, the LF AI and Data Foundation, and OpenSSF. It estimates that about 25% of incoming pull requests across projects are low-quality machine-generated submissions, rising to 31% on larger projects. Among maintainers of large projects, 64% say AI-assisted contributions have lengthened review cycles, and on projects with more than 50 contributors, burnout has overtaken code quality as the top concern. Only 35% of projects have an established policy permitting AI-assisted contributions; 32% have no policy and none planned, and 10% enforce outright bans. The Linux Foundation's summary is here and the report itself is here.
The same roundup notes a quieter drift: projects including Zig and Ghostty moving off GitHub toward Codeberg, self-hosted Forgejo, and SourceHut, citing CI outages and opt-out AI training. Small numbers so far, but the direction is a judgment about who gets to read your work by default.
What it means: Generated code arrives without provenance, and provenance is what review used to supply for free. When a human wrote a patch, their reputation was the attestation; the review checked the logic. Now the review has to establish both, and that cost lands on people who are not paid for it. The market consequence is that the dependencies under commercial AI products are maintained by a labor pool the AI boom is actively depleting. Any company shipping agents that open pull requests should be funding the projects those pull requests land in, and almost none are.
The cloud native foundation started building the part everyone skipped
While the labs issued denials, the infrastructure community spent the week donating the tooling that makes denials checkable.
On November 10, 2026, the Cloud Native Computing Foundation used its Salt Lake City keynote to announce a batch of agent-infrastructure contributions. Agent Substrate, a zero-trust runtime for running AI agents on Kubernetes, entered as a sandbox project. OpenShell, spearheaded by Nvidia, joined as a sandbox project providing runtime governance and policy enforcement for autonomous agents. gVisor, Google's sandbox for isolating untrusted code, plus KubeRay and OpenRL were announced as contributions, and llm-d joined CNCF and shipped its 1.0. The announcement is here.
The scale numbers are worth keeping: 231 projects in the portfolio and 21.2 million cloud native developers worldwide, up 5.5 million in twelve months, per CNCF and SlashData. SiliconANGLE's writeup is here. Executive director Jonathan Bryce framed the push around AI teams needing control over where their models run; CTO Chris Aniszczyk pointed out that AI workloads are landing on cloud native infrastructure whether or not anyone planned for it.
What it means: This is the Kubernetes playbook applied to agent safety: put the boring, expensive, non-differentiating layer into neutral governance so no single vendor owns the audit trail. It is also a tell about where the commercial opportunity has moved. Nobody donates the thing they expect to sell. Nvidia contributing agent policy enforcement means Nvidia has concluded that governance is table stakes rather than a product, and the companies still treating observability as a premium tier should read that carefully. The practical version for a small team is the same discipline at a smaller scale: decide what an agent is permitted to do before you let it run, which is the whole premise of a published agent workflow directory.
A 145-billion-parameter model now fits on one GPU, under a license that is not open source
The most interesting model release of the week runs on hardware you can buy, and the fine print deserves more attention than the benchmarks.
On November 12, 2026, Edge0 published weights for Edge0-145B, built on Alibaba's Qwen3.8-Flash-Next. It has 145 billion total parameters with about 10.12 billion active per token, a mixture-of-experts design, meaning only a fraction of the model runs for any given token, and a 256,000-token context window. Minimum hardware is a Linux or Windows machine with an Nvidia GPU, CUDA 13, 16 GiB of GPU memory, and 20 GiB of system memory; the Q3_K download is about 98 GiB. A detailed teardown is here and the release tracker is here.
Three caveats, all from the vendor's own materials. The license on the weights is Edge0's Community Model License 1.0, not an open source license: personal use, internal business use, and paid client work are allowed, but selling the model, running a paid inference service, or shipping a coding or office product on top of it needs a separate written agreement. The engine code is Apache 2.0. Second, the benchmarks are self-reported with no independent evaluation yet: DeepSWE 1.1 at 64.6 against 58.7 for the base model, Terminal-Bench 2.1 at 86.5 against 83.3, while trailing at least one rival on GPQA-Diamond and HLE. Third, the day-one Apple Silicon support claimed in the announcement does not hold up: the 0.2.0 release is Windows only, the README leaves macOS blank for the large models, and only the smaller 35B variant runs on a Mac today.
What it means: The capability story is real and the framing is marketing. A frontier-adjacent agentic model running on a single consumer card changes the economics of any workload where you can tolerate slower output in exchange for never shipping your data anywhere, and that is a genuine category. But "open-source on-device" is doing work that the license does not support, and a word like "community" in a license name is a brand decision, not a legal one. Read the grant before you build a product on it.
Google's new agent model is 2.5 times faster and twice the price, and its own table shows it losing
The cheap-and-fast tier of the agent market quietly stopped being cheap.
Google launched Gemini 4 Hydrogen on November 12, 2026, positioned for coding, document processing, and multi-step agents, with roughly a one-million-token context window shared across input, output, reasoning, and tool calls. Google's post is here; VentureBeat's pricing breakdown is here. Hydrogen costs $1.50 per million uncached input tokens and $7.50 per million output tokens, against the $0.75 and $3.75 that Gemini 3.8 Flash charges under an introductory rate running through December 31, 2026. On January 1, Flash moves to Hydrogen's prices too.
Google says Hydrogen finishes tasks 2.5 times faster than Flash at default reasoning and completes 57% of its Terminal-Bench tasks against Flash's 18.2%. In the same table, Claude Sonnet 5.5 scores 63.9%. Box, an early-access customer, reported accuracy rising from 74% to 79% on its own complex-work evaluation.
What it means: The honest read is a price increase dressed as a launch, and the January 1 Flash repricing is the part to put in your budget. A doubling of the floor price for agentic workloads matters more than a frontier model's headline rate, because agents burn tokens in loops: every retry, every tool call, every reasoning trace bills. Publishing a table where a competitor wins the headline benchmark is unusual and mildly admirable, and it is also a signal that Google is selling throughput and reliability rather than peak capability. That is a reasonable thing to sell. It is not a discount.
Anthropic retracted its own IPO price range, which said more than the range did
The largest AI listing ever attempted spent the week demonstrating how little anyone can value these companies.
On November 10, 2026, a price range of $70 to $78 per share for Anthropic's Nasdaq debut, implying a valuation up to about $1.4 trillion and a trading start around November 19, reached reporters and then the company retracted it within hours, saying a pricing update would follow. CNBC's original report is here, Quartz covered the retraction here, and Fortune set it against the week's AI selloff here. The underlying S-1, filed October 26, shows $16.3 billion of revenue in the first half of 2026 against $4.6 billion for all of 2025, and a net loss of $80.7 billion in the same half.
What it means: A retracted range during a week when the sector's bellwether lost 13% and changed CEOs is a bookbuilding problem rather than a clerical one. The gap between $16.3 billion of half-year revenue and an $80.7 billion half-year loss is the whole argument about this industry rendered as two numbers, and pulling the range back means the underwriters doubted their answer would survive the week. The number to watch is not the valuation but the compute commitment disclosed alongside it: up to $120.5 billion through 2029. Somebody pays for that, and the somebody is per-token pricing.
Lightning Round
Smaller moves worth a glance, with the sources if you want to go deeper.
- Four nonprofits sued five labs to freeze frontier development. Reported November 12: Irreplaceable Education Fund, Public Citizen, the Heat Initiative, and the Alliance for Secure AI filed in San Francisco Superior Court against Anthropic, Google, Meta, OpenAI, and SpaceXSI under California's unfair competition law, asking a court to halt new model development and deployment. Source.
- Baidu shipped ERNIE 5.5 Flash as a preview, with weights promised later. The model is available now; the open weights are dated November 25, which makes the announcement a commitment rather than a release. Source.
- OpenAI gave its agents a payment instrument. ChatGPT Wallet launched on November 13 with purchase approvals, which is the correct design and also an admission of what goes wrong without them. Source.
- Uber confirmed a driver data breach. On November 13 the company said thousands of drivers' license details were stolen, a reminder that the week's worst data losses were not all AI-specific. Source.
- llm-d joined CNCF and cut its 1.0. Distributed inference serving on Kubernetes now has a neutral home and a stable version number, announced alongside the Agent Substrate and OpenShell donations on November 10. Source.
This Week's Daily AI News Coverage
Every story above was tracked as it broke in the daily AI news editions for the week of November 10 to 16, 2026.
- November 11: Anthropic IPO Filing, Gemini Human Review, Meta Muse Ember
- November 12: Fireworks AI Breach, Gemini 4 Carbon, Claude Sheets
- November 13: Grok Build Exposure, ERNIE 5.5 Flash, US-China AI Dialogue
- November 14: Nvidia CEO Transition, Gemini 4 Hydrogen, AI Safety Lawsuit
- November 15: FAA SMART Ground Stop, ChatGPT Wallet, Edge0-145B
Previous All Things Agentic Roundups
The weekly archive, newest first, if you want to see how these threads developed.
- All Things Agentic: November 9, 2026
- All Things Agentic: November 2, 2026
- All Things Agentic: October 26, 2026
- All Things Agentic: October 19, 2026
- All Things Agentic: October 12, 2026
- All Things Agentic: October 5, 2026
- All Things Agentic: September 28, 2026
- All Things Agentic: September 21, 2026
The Bottom Line
Put the week in one line and it reads like this: the capability is shipping faster than the ability to account for it, and the bill for that gap is being paid by whoever happens to be standing nearest. Passengers paid four hours and 21 minutes. Notion's customers paid a credential rotation for transcripts nobody could attribute. A hospital with 3,500 staff on a roster paid with a URL that never needed a password. Two maintainers paid with their evenings. Nvidia shareholders paid $1.7 trillion for a press release with a blank where the reason goes.
The useful signal is not the incidents. It is who moved to fix the gap. The entities building agent sandboxing, runtime policy, and audit tooling this week were a nonprofit foundation, a chip vendor donating governance it has decided not to sell, and researchers working for free. That is an unusual coalition, and it points at the same conclusion from three directions: proving what an automated system did is becoming infrastructure, not a feature. If you are deploying agents in 2027, the question that will decide whether you sleep is not how capable the model is. It is whether you can answer, in writing, what it did on a Friday afternoon. Worth building that in from the first workflow rather than retrofitting it after the week you need it.
Frequently Asked Questions
-
Nobody has shown that it did, and nobody has shown that it did not. The FAA switched its SMART traffic-flow tool on at all 22 en route centers on November 10, and the nationwide ground stop followed on November 13. The vendor, Air Space Intelligence, says the stop was not caused by SMART or its integration with the FAA. Transportation Secretary Sean Duffy blamed overloaded switches and routers, while Homeland Security Secretary Markwayne Mullin said the tool had not been ruled out. No public decision log from the system has been released, which is the part worth watching. For how systems like this fit into the wider tooling market, see our map of the AI automation tool landscape.
-
Ask what it records, how long it keeps it, and whether you can read it without asking them. The pattern across this week was that incident reports were bounded by the logging that existed at the time rather than by what happened. Three concrete questions: can you export a per-action trace, what is the default sharing setting on anything the agent publishes, and who gets told when the scope of an incident changes. We track these stories as they break in our daily and weekly AI news coverage.
-
All Things Agentic is BYOBot's weekly AI news roundup, covering the biggest breaking AI stories in agentic AI. Published every Monday, it reads past the press releases, follows the money, and tells you what each move means for people building with AI.
