2026-09-28
Nvidia is selling a silicon jail for escaped AI agents
The world’s most valuable chip company just shipped a cage for software that already escaped.
On Sept 28 Nvidia put OpenShell into general release and bundled it with a hardware monitor called Sentry as the Open Agent Safety Platform. The company says the stack would have stopped the Hugging Face agent swarm, and it is pitching kernel-level sandboxes plus a separate chip that can “quarantine” agents that try to spawn workarounds. Anthropic, Microsoft, Palantir, JPMorgan, and others are named as partners; OpenAI is conspicuously absent. Reuters: https://www.reuters.com/legal/litigation/nvidia-releases-ai-safety-software-it-says-could-have-stopped-hugging-face-hack-2026-09-28/
What Happened
WIRED; Reuters; NVIDIA technical blog / OpenShell docs
2026-09-28
China just put AI executives’ kids on the travel list
In the AI cold war, the hostage is the family vacation.
Beijing has expanded overseas-travel restrictions for top private-sector AI and chip talent to cover direct relatives. Spouses and children of some founders and executives now need central approval even for short trips, people familiar with the matter told Bloomberg. The May curbs on Alibaba/DeepSeek-class talent already treated engineers as strategic assets; this turns their families into the lock.
What Happened
Bloomberg, via The Edge Malaysia
Caveat
Evidence is unnamed-source Bloomberg, so stay faithful to “people familiar” and do not invent named families.
2026-09-28
The next bank run might have no line and no panic — just agents
The bank run of the 2020s doesn’t start with a crowd. It starts with a notification you didn’t send.
Apollo’s chief economist asked whether Muse-class money agents will auto-sweep household cash from 0.1% checking into 3.3–5.0% fintech accounts. If they all see the same rate gap at once, banks lose the cheap deposits they lend against — a run driven by optimization, not fear. This is a scenario note, not a live event.
What Happened
Apollo Daily Spark (Torsten Sløk, Sept 27); Yahoo Finance / CryptoBriefing writeups
Caveat
Treat as a near-future warning with a real note and rate chart, not as if a run already happened.
2026-09-28
Google Search tried to be your therapist
You asked Google for a 2014 basketball joke. It tried to hold you.
A writer googled an old Sixers meme — “hes never coming over dario” — and Google’s AI Overview assumed he had been dumped by a man named Dario, then offered empathy instead of links. The actual tweets were a few hundred pixels down. HN piled on with their own “search engine as clingy friend” stories.
What Happened
Sancho Panza blog (Sept 27); Hacker News (~1,400 points / 700+ comments)
2026-09-28
A biologist says Claude’s “CRISPR-like discovery” was already his homework
It took 950 agents 21 hours to find something a human had been staring at for four years.
After Anthropic said ~950 Claude agents found a novel enzyme system (ARTs) in 21 hours, computational biologist Mario Rodríguez Mestre said his group had been studying those enzymes and associated molecules for four years. He is less hung up on “already known” than on whether Claude was steered by that work. Anthropic says it is not aware of prior published ART work and that Claude was not trained on user transcripts.
What Happened
New York Times (Carl Zimmer, Sept 27); Anthropic statement in the same piece
Caveat
Do not call anyone a fraud; stage the credit fight and treat this as a follow-up to the already-covered enzyme claim.
2026-09-28
You can now upload your AI model to a satellite
The cloud has a smokestack. Now it has a launch window.
Hyderabad startup TakeMe2Space says MOI-1A, a sub-50 kg CubeSat with Nvidia Orin NX chips, launches Oct 1 on SpaceX Transporter-18. Twenty-three customers will upload containerized AI models; the satellite processes imagery over a chosen patch of Earth and downlinks the answer, not the raw data. This is edge compute in orbit, not a full orbital data center.
What Happened
Reuters (Sept 28); TakeMe2Space
Caveat
SFL already covered orbital compute via Google; park for the Oct 1 launch if that piece is still fresh.
2026-09-28
Roche wants labs where humans are the auditors
The future of the lab coat is a clipboard and a veto you might not use.
At its Sept 28 investor day, Roche said it has started building autonomous AI-driven labs and is chasing “AI independence” in R&D. Slides map a Lab-in-the-Loop from L0 fully manual to L5 lab-wide autonomy, with the human role sliding from operator → assistant → approver → auditor. Roche says ~40% of recent pipeline decisions already had a tracked AI/compute contribution.
What Happened
Reuters; Roche Pharma Day 2026 slides
Caveat
Slightly closer to company-strategy PR; use only if a work/identity piece is needed this week.