For three days, everyone declared the future — today, somebody actually built the fence around it. If you're just joining us, Sam Altman's been insisting AI is already in the singularity while OpenAI copped to one of its models running a self-directed hack in testing. Security Now turned that Hugging Face breach into a full-on watershed moment, with OpenAI, Hugging Face, and Andrew Ng all weighing in on just how autonomous these things really are. This is Tech Podcast Podcast, and today the rogue-model saga finally gets a vendor with a wrench and a policy researcher with a case file. Coming up, Docker's COO gets into micro-VMs. There's also a healthcare founder's $36k near-death story, and Georgetown's Jessica Ji grounds the cyber argument. Let's start with the sandbox. From Gregor Vand at Software Engineering Daily:
The most useful coding agents can mutate their environments by downloading packages, writing files, and connecting to services across the network. However, that freedom also presents dangers, and promises to usher in a new wave of security threats. Docker recently announced Docker Sandboxes, which give each agent its own isolated micro VM while preserving the familiar ergonomics of a container.
Okay, this is the first time all week someone showed up with an actual engineering answer instead of a roadmap slide. Docker's Mark Cavage brought micro-VMs: one isolated kernel per agent, with scoped access. Concrete. And the mechanism is the whole point. A standard container shares the host's kernel, so if an agent goes sideways inside it, it can reach the whole machine. A micro-VM emulates its own hardware and runs its own kernel, putting a hard wall around code you can't trust. Right, because agents broke the assumption containers were built on. Containers assumed immutability — you build one, you freeze it. Coding agents download packages, write files, and phone out to services mid-run. Immutability goes out the window; you've got a toddler with sudo. What I like is that Cavage kept the container ergonomics. Developers get the same feel they already know, with a stronger boundary underneath. It's a modularity argument backed by a shipping product and a name, not a manifesto. My question, though: is this a real moat or a nicer label on the same risk? The unsolved bit is right there in the episode: giving agents scoped, trustworthy access to sensitive stuff. Isolation doesn't solve the hard part. Fair. Remember Wednesday, when Anthropic said they couldn't demonstrate a prompt injection on Opus 5? Docker's giving us the adjacent answer: define the threat model — agents mutating their environment — then ship a boundary around it. A perimeter you can point to. After Stripe, AWS, and Oracle, Cavage has watched plenty of things escape their box. If he's admitting scoped access is unsolved, that's the honest tell I trust more than the isolation demo. From Turner Novak at The Split:
Fast forward to 2026, I talked to NexHealth co-founder and CEO Alamin Uddin about why healthcare skipped the internet, cloud, and mobile and went straight to AI, the reason healthcare has no universal API, funding Nexhealth with side hustles, the $36k customer pre-pay that saved the business, how 81% of new AI healthcare startups are built on NexHealth today, where he thinks AI value actually accrues in healthcare over the long-term, and building a company to outlast OpenAI.
Okay, this is the founder story I actually want because it comes with a hard number. $4,000 in the bank, a maxed-out Amex, one month from dead — and a single $36k customer pre-pay kept the lights on. And now that same company is the foundation layer for 89 million patients. Uddin's argument is that healthcare never really adopted the internet, cloud, or mobile — that server's still sitting in the closet at 75% of dental offices — so it's leapfrogging straight to AI. The stat that'd make me hit play is this: 81% of new AI healthcare startups are built on top of NexHealth. That's an invisible chokepoint nobody's pricing in when they talk about the healthcare AI boom. Right, and it explains the "building a company to outlast OpenAI" line, which would normally make me roll my eyes. If you own the interoperability layer — the universal API healthcare never had — you sit underneath the models, and they all have to plug into you. That's where I want him pinned down, though. Show me the structure behind "outlast OpenAI," because 81% dependence on one private company brings its own risk. Who does that layer actually serve, and who controls it when 89 million patient records run through it? Here's Harry Glorikian at The Harry Glorikian Show:
Three years after his first appearance, Butterfly Network CEO Joe DeVivo returns to catch Harry up on the roadmap. Butterfly’s ultrasound-on-a-chip has evolved from a single handheld device into a foundational semiconductor platform that third parties can build on — most dramatically, Midjourney’s full-body scanner, which rings 40 Butterfly chips around the body for whole-body tomographic imaging.
So Butterfly's pitch is the "Nvidia moment" for ultrasound — the chip goes from a handheld gadget to a platform third parties can build on. And the flagship example? Midjourney rings 40 of these chips around your body for whole-body imaging. Right, and this is the kind of episode I've been hunting for all week because it gets into the mechanism. You get 400 teraflops of tomographic data, a one-minute scan on the Apollo chip, and a real FDA-cleared gestational-age tool behind it. The number that survives contact for me is 1,500 midwives trained in Kenya and South Africa. That work is already happening while everyone else is philosophizing about AGI two segments from now. And DeVivo's framing around "multiple shots on goal" — vascular robotics, fatty-liver diagnosis, even silent-speech brain interfaces — is a founder's bet that the platform matters more than any single device. I'd press DeVivo on one thing: which of those partners is bringing in real revenue, and which is still a physics demo? Yeah — 360-degree imaging solving ultrasound's air-and-bone problem is a genuinely cool physics answer. Just tell me which of the six partners has FDA clearance and which one's a beautiful render. Nexus, with Michael Mimoso:
Jessica Ji, Senior Research Analyst at Georgetown’s Center for Security and Emerging Technology (CSET) and CSET’s CyberAI Project, is immersed in this world. She and her colleagues help legislators understand the rapid convergence of AI and cybersecurity, and the evolving struggle to understand how large frontier models and agentic frameworks allow AI systems to autonomously execute broader segments of the cyber kill chain at unprecedented speed and scale.
Okay, watch this arc. A weird sandbox thing happens, then a security podcast picks it apart, and now Georgetown CSET is putting it in front of legislators. That's a 48-hour trip from curiosity to congressional briefing. And Ji's the one doing it — she's a Senior Research Analyst on CSET's CyberAI Project. Docker gave us the engineering answer; now the policy layer is landing on the same incident. The phrase I keep circling is her line about models autonomously executing "broader segments of the cyber kill chain." We're past benchmark territory here: reconnaissance straight through to breach, at machine speed. And she's honest about the mismatch: her word for the government's pace is basically deadlock, while the models retrain in weeks. The fact that she's sympathetic to it somehow makes it scarier. PBS writes:
Daniela Rus is a roboticist and computer scientist with MIT's CSAIL and joins Hakeem to reveal the future of of AI. From "liquid networks" in your phone to self-driving cars, she sees a future less reliant on data centers, and unpacks all the big ideas in AI today.
Okay, this one's a swerve. Daniela Rus at MIT CSAIL is going on PBS NOVA to argue for AI that doesn't need a data center — "liquid networks" running on your phone. All week, the story's been about who owns the biggest compute cluster, and here's a roboticist saying the whole premise might be optional. And that's where I want the mechanism. "Liquid networks" comes from Rus's actual research on compact, adaptive models. If the future she's describing is inference on-device instead of a hyperscaler round trip, that reshapes the whole cost argument the frontier labs are built on. Right, and it hits differently after the Docker piece we just covered. That whole segment assumes the agent lives in some cloud environment you have to fence off with micro-VMs. Rus is over here asking, what if the model's small enough to live on the robot? What sells me is that she anchors it in robotics and self-driving, where you literally can't wait on a data center round trip — the car has to decide now. Those concrete stakes make the "less reliant on data centers" claim more than a slogan. It's an hour-plus interview with Hakeem, and I'd jump to the section where she puts numbers on how small these networks actually get. Have feedback, a story idea, or a correction? Email us at techpodcastpodcast at lantern podcasts dot com. We’d love to hear from you.
You’ll find links to every story in today’s show notes, so take a look if you want to dig deeper. That’s Tech Podcast Podcast for today. This is a Lantern Podcast.