Podcast · Monday, November 9, 2026
The AI trade now depends on one person's judgment and on borrowed money, and both were questioned in the same weekend.
Today's headline
The AI trade now leans on one person's judgment and on borrowed money, and both were questioned this weekend. Nvidia's board opened an investigation into its CEO. CoreWeave's bonds now yield more than sixteen percent. For builders, the practical news is lower down: a live OpenSSH bypass, new Google chips, and a stark look at what one attacker did with local AI agents.
On the clock
GitHub is running a test today. Between sixteen hundred and eighteen hundred UTC, workflows that use the pull request target trigger in public repositories will fail if the new default rule would block them. The real block starts December second. Check your Actions insights now, and fix or allow those workflows before the next test on November sixteenth.
Tomorrow is the federal patch deadline for the Brotli flaw known as Node2Shell, and Node dot js ships security releases on Wednesday. Patch anything that bundles a Brotli decoder, and reserve a window for Wednesday.
GitHub Actions and webhooks were also degraded this morning. Run starts were delayed and some webhook deliveries failed, then service recovered. GitHub says a root cause analysis is coming. Check for deploys that never triggered.
Security
Attackers are now exploiting SSHole, the OpenSSH password bypass. Severity is high, and it is exploited in the wild. It affects Linux systems built on musl, such as Alpine, Buildroot, OpenWrt and Yocto. A very long password can log in to accounts that have no password, root included. So far only botnets and crypto miners are using it. Upgrade OpenSSH to ten point six p two, or turn off password login, and then check your logs for unusually long passwords.
Acronis described a single attacker in Taiwan who ran twelve AI agents against emergency services and other targets in at least eleven countries. Commercial models mostly refused. So he ran the agents on local models with their safety training removed, including versions of DeepSeek, GLM and Qwen. They pulled dispatch records and hundreds of thousands of phone records, mostly through default passwords and SQL injection. There is no CVE here, so your scanners will be silent. The agents also invented flaws, and more than half of one scanner's top findings were wrong. Log default-credential logins on your admin panels.
Two smaller items. Google DeepMind says its agents cut the cost of forging signatures with 1024-bit RSA keys by more than half. It only matters for old signing hardware. If you still run it, move to larger keys. And PyPI added a Security tab that shows whether a release has build provenance and warns when it suddenly changes or loses it. Missing provenance does not prove a package is malicious, but a sudden change is a reason to look at the maintainer's history. Use it on your top dependencies.
Agents
Perplexity described how its Teammates product keeps agent work durable. Each project is a Git repository. Agents work in separate sandboxes and submit changes back. Perplexity says fewer than two percent of submissions are rejected for stale state. That is the company's own figure, but the design is worth copying.
Cursor added voice calls with its agents, running on Grok Voice. You can ask what changed in a diff while work continues. Privacy mode gives zero data retention.
Simon Willison suggests hiding agent-written pull request descriptions inside a collapsible section, with a short human summary on top. It is a cheap fix for review fatigue.
Models and pricing
Google added two systems to its eighth-generation TPU line. The 8tx keeps data and model weights encrypted while they are in use, aimed at regulated workloads that want to fine-tune and train agents without moving sensitive data. The 8i Max puts two chips in one package, doubling compute and memory in the same rack. It is Google's fourth TPU system this year, and PyTorch, vLLM and SGLang are supported. Google also open-sourced TPU Sync, which it says cuts cache transfer latency by more than ninety percent. That is a vendor-reported figure.
Pinterest replaced a multistage clustering pipeline with one fine-tuned four-billion-parameter model. The company reports email clicks up one point one percent. The model reads up to one hundred fifty user actions over a year and names the goal behind them. It is a modest, company-reported gain, but from a much simpler system, which is the point for your own pipelines.
Business and policy
Nvidia's board opened an independent investigation after CEO Jensen Huang admitted a relationship with an executive at customer Megaspeed. US intelligence agencies reportedly call her a Chinese informant. The board says it has no reason to believe secrets were compromised, and China's foreign ministry calls the claim unsubstantiated. Nvidia shares fell about five percent before the open. Incoming Democratic committee leaders promise oversight, which weakens Huang's lobbying against tighter export controls.
CoreWeave reports earnings tomorrow, and Nebius on Thursday. CoreWeave's bonds due in twenty thirty-one yield more than sixteen percent. If you rent neocloud capacity, ask your provider about customer concentration and contract terms.
SemiAnalysis says Envision built the first hundred twenty megawatts of a planned eight hundred megawatt AI campus in Inner Mongolia in under a year, at about a fifth of comparable US costs.
IEEE Spectrum looked at how the US could track advanced AI chips. The leading idea is ping-based location checks, which time round trips to trusted servers. It can only catch diversion after it happens. It misses chips that are offline, and a thief with stolen keys could fake the signal. So physical audits stay necessary. A House committee has passed a bill that would require verification, and the Nvidia investigation raises its odds.
Chinese labs are recruiting in San Francisco. MiniMax lists about sixteen openings, including a research lead for its core model, and Moonshot wants US solutions architects. Rest of World says poaching top talent from American labs will not be easy.
Before you go
If any host you run uses Alpine or another musl-based system, upgrade its SSH server today. This exploit needs no valid password, and attackers are already trying it.
Action items
Upgrade OpenSSH to ten point six p two on musl-based hosts today. Audit every public repo that uses the pull request target trigger before December second. Check your top Python dependencies in PyPI's new Security tab. Move any 1024-bit RSA signing setup to larger keys this quarter. Add local, safety-stripped models to your agent threat model, and log default-credential logins on internal admin panels.