Podcast · Thursday, November 12, 2026
The AI stack's weakest points today are people and credentials, not models.
Today's headline
The weakest points in the AI stack today are people and credentials, not models. Nvidia's board met overnight on whether to remove its chief executive, and six analysts cut the stock. Separately, thousands of hijacked GitHub accounts planted a secret stealer in repositories, and a critical flaw landed in the web server that runs much of the internet.
On the clock
Alibaba's Qwen4-Max preview is available only on its Token Plan, and only until November twenty-fifth. It offers a one million token context window. Finish any evaluation before the cutoff.
Next week brings several dates. GitHub retires three Copilot models on the sixteenth, and Gemini retires several text-to-speech and live preview models on the seventeenth. Repoint your model pins now.
Security
An attacker used stolen credentials from nearly five thousand GitHub accounts to plant a secret stealer in repositories those accounts could write to. Security firm Aikido says it ran in at least eighteen hundred repositories and took Actions secrets, cloud keys and AI tool tokens. The attacker deletes the run logs within a minute. There is no patch, and scanners will be silent. Search your organizations for pull requests titled "docs: dependency audit notes," and rotate every secret in any repository where one appears.
Apache HTTP Server has a critical remote code execution flaw in its HTTP/2 module. It needs no login, and it is rated nine point eight. Wiz says no exploit is known to be in use. Red Hat says public research shows one, so treat the risk as live. Upgrade to version 2.4.71, or your distribution's patched package, on every host with HTTP/2 turned on.
MongoDB fixed a critical flaw that lets an unauthenticated attacker crash a replica set member. The crash happens when that member later becomes primary. MongoDB says it is not aware of attacks. Upgrade to the fixed release for your branch.
Node.js fixed sixteen flaws. Two HTTP/2 bugs can crash servers or leak memory. A third can leak TLS private keys from certain HTTPS servers. Move to twenty-two point twenty-three point five, twenty-four point twenty-two point three, or twenty-six point thirteen point one.
N-able's N-central is in trouble again. A researcher published new exploits that work against the build meant to fix the last batch. N-able says they need a non-default Legacy API exposed to the internet. Turn it off today.
ConnectWise says a campaign hit at least two hundred eighty-eight endpoints across forty-eight organizations through ScreenConnect. It disabled five third-party extensions and says it found no flaw in ScreenConnect itself. Check whether your instances run those extensions.
Red Hat Quay has a critical flaw rated nine point eight. If its signing-key service is reachable by untrusted clients, an anonymous attacker can take over token signing and read or overwrite any repository. Apply Red Hat's advisory, and keep that service off untrusted networks.
Notion confirmed that the Fireworks AI breach exposed some Notion AI chat transcripts. Fewer than one percent of workspaces were affected, and most transcripts cannot be tied to a customer.
Darktrace showed that a stolen Claude API key can turn Claude Code into a remote-controlled implant in a lab test. It is not a vulnerability, and Darktrace has not seen it in the wild. Protect those keys like any other credential.
Agents
AWS released an emulator that runs its services on a laptop, and agents built all of it. They will maintain it too. AWS released it into the public domain because it was unsure agent-written code can be copyrighted. Point your tests at it, and agents never need real cloud credentials.
Microsoft's research team scanned parts of the Linux kernel with an AI pipeline. It produced fifty-nine public bug reports and thirty-three CVEs. The raw scan produced more than twenty-six thousand findings, and sorting them took months. Plan for that triage cost if you adopt the same approach.
DeepMind says its AlphaProof Code verified the core of Signal's encryption in four weeks. Agents wrote six million lines of proof, and humans reviewed about one percent. This is DeepMind's own result.
Models and pricing
Google launched the Gemini Decisions API. One call scores up to one thousand questions in a single pass, in about a quarter of a second. It costs ten cents per million input tokens, and output is free. Use it for routing, policy checks and tool selection. Google's benchmark claims are its own.
Google also split its latest chip family. One new system targets confidential inference. Another pairs two inference chips for long-context agents. Google open-sourced a library that cuts cache transfer time, in its own tests. The interfaces are changing quickly, so pin a commit.
Business and policy
Nvidia's board met Wednesday on whether to remove Jensen Huang. The New York Times reported a second relationship, with a direct report. By Thursday morning, no decision had been announced. Six analysts downgraded the stock, and it fell about eight percent before the open. Treat anything built on a single Nvidia roadmap as carrying new risk until the board speaks.
The Information reports Commerce is finalizing a rule against Chinese firms renting advanced chips through data centers abroad. It could extend to seven more countries, including Malaysia, Singapore and Thailand. It could land this month. If you run cloud regions in Asia, expect new customer screening.
Anthropic told investors that its own model ported Claude to AMD chips in three days instead of months. That is a company claim, offered as evidence that Nvidia's software lock-in is weakening.
Fractile, a British chip startup, raised seven hundred twenty million dollars. Australia proposed mandatory safety audits for frontier AI labs. That is a proposal, not law.
Action items
Search every GitHub organization you own for the stealer's pull requests, and rotate secrets today. Upgrade Apache HTTP Server to 2.4.71 on every host with HTTP/2 enabled this week. Move Node.js and MongoDB to their fixed releases. Disable the N-central Legacy API, and audit ScreenConnect extensions before your next maintenance window. Rotate Notion AI and Fireworks-linked credentials set up before October twenty-fourth, by month's end. Finish Qwen4-Max evaluations before November twenty-fifth.
Before you go
Open your GitHub audit log before lunch and look for pull requests opened by people who never opened one before. A stolen token looks like a normal teammate, right up until it is not.
Here is the short list. Rotate Actions secrets in any affected repository today. Patch Apache by week's end. Update Node.js and MongoDB. Turn off the N-central Legacy API. Finish Qwen4-Max testing by the twenty-fifth.