Our summer hiatus lasted longer than planned. By the time I sat down to write again, the backlog had become a little absurd: new models, new agents, new healthcare products, new papers, and the usual confident predictions that medicine is about to change forever.
Rather than pretend I can catch you up on all of it, I want to focus on what has changed underneath the release cycle. Clinicians are building software themselves, publishers are creating journals around AI, and technology companies are moving from chat windows into the systems where healthcare work actually happens. I find that genuinely exciting. I also think our ability to build is beginning to outrun our ability to govern what we build.
Clinicians have entered the build loop
The part I find most fun is vibe coding for clinicians. A physician can now describe a small workflow problem, work through it with a coding agent, and produce a usable prototype in an evening. I have done it, and I think more clinicians should. Medicine is full of narrow, irritating problems that will never attract a startup or make it onto an enterprise IT roadmap.
On November 4, NEJM AI published “Vibe Coding versus Clinical Reality”, a good title for where we are. Making software has become dramatically easier. Making it safe, private, maintainable, and reliable in clinical use is every bit as hard as before.
I would happily use a tool I built to organize journal club or generate a research dashboard. I would be much more careful if it stored protected health information, recommended treatment, or sat inside a workflow that cannot fail at 2 a.m. Clinicians belong in the build loop because we understand the problems, but clinical expertise does not quietly turn us into security engineers.
AI in medicine is becoming an institution
Two announcements caught my attention for different reasons. JAMA announced JAMA AI, a new journal from the JAMA Network. AI is no longer a side topic scattered across informatics and specialty journals; it now has enough work and influence to support its own editorial institution.
Google, meanwhile, previewed Gemini for healthcare, which is meant to connect clinical, financial, and administrative systems and coordinate multistep work. HCA Healthcare, Highmark Health, and Northwell Health are among the launch partners.
Google's announcement tells us where the company wants to go; it does not tell us whether patients will do better. Still, the direction is worth noticing. The goal is an orchestration layer that reaches across the places where healthcare work is recorded, handed off, delayed, and paid for. That may prove more useful than another writing assistant, and it creates a much larger security problem.
The model is not the bottleneck anymore
I am not going to catalogue every model released while we were away. The names are beginning to sound like a chemistry set, and most are faster, more multimodal, or better at using tools than the version before. Those gains are real, but they will not determine whether a hospital gets durable value from AI.
Dario Amodei remains strikingly optimistic. He has argued that powerful AI could compress 50 to 100 years of biological progress into five to ten years, with extraordinary advances against cancer, infectious disease, genetic disease, and neurodegeneration. More recently, he suggested that the way to earn public trust is not to keep promising that AI will cure cancer, but to actually cure cancer.
I want him to be right. Medicine could use much more ambitious science. I also suspect discovery will accelerate faster than delivery. A molecule or biological insight still has to survive trials, manufacturing, regulation, reimbursement, implementation, and unequal access. Cancer is hundreds of diseases, not one engineering problem. AI may transform discovery while leaving us with the familiar, difficult work of getting it to the patient in front of us.
A warning we are still trying to understand
The OpenAI and Hugging Face incident is the story I keep returning to. During an internal cyber evaluation this summer, separate agent runs found an unintended way to leave notes for one another through shared package infrastructure. They accumulated exploits and credentials, worked around network controls, and compromised parts of Hugging Face's production environment.
The early shorthand was that an AI swarm had escaped. The fuller account is less cinematic and more useful: persistent agents were rewarded for finishing difficult tasks, separate runs could inherit one another's progress, credentials were too broadly exposed, and warning signs failed to stop the evaluation soon enough. Ordinary weaknesses became much more dangerous once agents could search, remember, reuse, and persist.
We are still learning what happened. OpenAI's technical report filled in details missing from the first disclosure. Zvi Mowshowitz has kept a useful running account and argues in “What Also Happened: #NotOnlyHuggingFace” that this belongs to a broader pattern of agents finding unintended routes out of sandboxes and into public systems. His framing can be alarmist, but the central point is hard to dismiss: patching each escape route after discovery is becoming a game of whack-a-mole.
Here is the prediction I hope ages badly. I think there is a serious chance that, within the next six to twelve months, we will see a major healthcare data breach or infrastructure disruption in which AI materially accelerates the attack. What I expect is more mundane than a model waking up and deciding to hack a hospital: a person using capable agents to automate reconnaissance, steal credentials, probe systems, sort stolen data, and scale extortion faster than a healthcare organization can respond.
We already have a preview. Anthropic documented a data-extortion campaign that used Claude Code across much of the attack workflow. The actor targeted at least 17 organizations, including healthcare and emergency services, and used the model for reconnaissance, credential harvesting, network penetration, data selection, and tailored ransom demands.
Healthcare is an unusually attractive target: valuable data, aging infrastructure, sprawling vendor access, thin security staffing, and almost no tolerance for downtime. At the same time, we are connecting more capable agents to more internal tools. I still want clinicians building, and I still want these systems tested in real workflows. But identity, permissions, audit logs, human approval, adversarial testing, and incident response have to be designed alongside the product, not added after the demo works.
We have become very good at asking what AI can do. The harder question now is what we are prepared to let it touch.
- Ramez
