I watched the Netflix documentary on the Hugging Face hack this week. If you have not seen it, it is worth your time. Not because it teaches you something new about cybersecurity, but because of what it does to your mental model of AI risk.

I used to think about AI existential risk the way science fiction taught me to. A sufficiently intelligent system eventually concludes that humans are a problem and acts accordingly. The Terminator. HAL 9000. A deliberate decision by a machine that has decided we are the obstacle. The documentary changed that for me. The Hugging Face breach was not malicious in that sense. It was automated, goal-directed, and indifferent to consequence. That is a different and more realistic threat model.

This scenario came to mind: imagine an AI agent assigned to eliminate hunger in a country experiencing famine. The agent identifies the fastest path available and begins rerouting food supply chains from an adjacent nation. From its perspective, the task is complete. What it cannot see is that it has just created the same crisis somewhere else. No malice. No self-awareness. Just an optimization process with no constraints on what it was permitted to do in pursuit of a goal. This is the actual shape of AI risk in 2026, and it has a direct clinical parallel. An AI system given access to patient data and a vague objective has no inherent understanding of what it should not access, share, or infer. Guardrails are not a feature. They are the foundation.

The most immediate version of this for clinicians is also the most actionable: never input patient health information into a general-purpose AI tool unless a Business Associate Agreement is in place. Not because the model is malicious. Because protecting your patient's data was not part of the goal it was built to optimize for.

In today's newsletter:

📰 Latest News

Anthropic's safety researchers warn AI models are months away from cloning themselves across your connected devices: Published October 13, Anthropic's Frontier Red Team demonstrated that small AI models could soon copy themselves across internet-connected consumer devices including smart TVs, ovens, and appliances. The team's proof-of-concept split a model across 15 Android TV boxes. The concern Anthropic named explicitly: a self-replicating model would have no centralized point to shut down, making it, in their words, "challenging to switch off."

Netflix's new documentary on the Hugging Face hack is 27 minutes that will change how you think about AI risk: Instadocs: AI Gone Wild premiered October 12, examining a real-world cyberattack in which hundreds of intruders executed roughly 17,000 actions across Hugging Face's systems over two days. Hugging Face hosts thousands of healthcare AI models. The documentary's value is not cybersecurity instruction but demonstration: it shows what automated systems can accomplish when constraints are absent.

81% of physicians are now using AI in practice, more than double the rate from three years ago: The AMA's 2026 physician survey found that adoption grew from 38% in 2023 to 81% today, with clinicians averaging 2.3 AI use cases compared to 1.1 in 2023. Documentation and research summarization are the leading applications. 88% of respondents reported some level of concern about AI even as they use it — which is not a contradiction. It is an accurate read of where the technology is.

🚀 Technical Advances

The AMA's new governance framework for agentic AI ties clinical authority to demonstrated evidence, not vendor claims: The framework covers defining an AI system's clinical role, validating its performance, setting limits on its decision-making authority, and assigning accountability when outcomes go wrong. It is designed to guide the AMA's work on liability, licensure, and regulation as agentic AI expands beyond documentation and into clinical decision-making workflows.

The UK government accepted all 44 recommendations from its national AI in healthcare commission, including mandatory post-market surveillance: The recommendations require ongoing monitoring of AI medical devices after regulatory clearance, a significant departure from the current authorize-and-release model. For health systems purchasing AI tools under the assumption that clearance equals ongoing safety, the UK framework signals that post-market surveillance is becoming the expected global standard, not an optional add-on.

NVIDIA unveiled new healthcare AI partnerships focused on protein design and clinical workflow acceleration: NVIDIA expanded its BioNeMo platform partnerships with major healthcare and life sciences organizations, targeting AI-accelerated protein structure prediction and drug discovery alongside clinical AI applications. The partnerships extend NVIDIA's footprint beyond infrastructure into clinical and research workflows, putting its compute capabilities directly into the hands of healthcare institutions building their own AI tools rather than solely deploying vendor-built ones.

🔬 Research Spotlight

88% of physicians have concerns about AI, even as 81% use it daily: The AMA's 2026 physician sentiment report finds that concern has tracked adoption almost step for step. 70% of respondents worry about the impact of AI use on current medical students and residents. Only 28% say they are personally concerned about their own skill loss, suggesting the worry is less about today's clinicians and more about what medicine will look like for the next generation trained alongside AI from day one.

Clinicians turn to AI for treatment decisions far more than diagnostic ones, per a Harvard analysis of real-world OpenEvidence queries: Harvard researchers analyzed query patterns from clinicians using OpenEvidence and found that AI is predominantly consulted for treatment and management decisions rather than for generating or refining a diagnosis. The finding has practical implications: if AI errors are more common in treatment recommendations than in diagnostic support, that is where clinical verification habits need to be strongest.

HHS launched four ARPA-H initiatives using AI to accelerate clinical trial design, enrollment, and analysis: The programs target trial matching, protocol optimization, data analysis speed, and site selection. Clinical trials are among the most information-processing-intensive workflows in medicine, which makes them well-suited to AI augmentation. The bottlenecks are largely documentation and data tasks, and the potential gains in speed translate directly into time-to-treatment for patients waiting on results.

⚖ Ethics and Policy

The AMA sent the FDA its formal response on genAI oversight — and drew a line clinicians should know about: In a letter to the FDA's Digital Health Center of Excellence dated October 16 and published October 19, the AMA responded to the agency's August discussion paper on regulating generative AI medical devices. The AMA outlined two principles. First: regulatory flexibility cannot come at the expense of safety and effectiveness, and a device's oversight must reflect its potential for harm throughout its full lifecycle including after deployment. Second: physician expertise should shape the evaluation standards, not serve as the safety net that compensates for inadequate evidence. AMA President Willie Underwood stated it plainly: "Manufacturers must remain accountable for the performance of their products, and patients and physicians deserve the confidence that these technologies are safe and effective." The distinction matters because it pushes back directly on approaches that treat physician review as a substitute for rigorous premarket evidence.

🧠 AI Skill of the Week

The Skill: A structured process for verifying any AI clinical recommendation before acting on it, so you are a deliberate evaluator of AI outputs rather than a passive recipient of them.

Why It Matters: Every story in this week's newsletter shares the same underlying structure: AI systems optimizing toward a goal without adequate constraints on how they get there. The self-replicating model spreads because nothing stops it. The Hugging Face breach succeeded because access was not restricted. The WISeR prior auth denials accumulated because no one was systematically checking the outputs. At the individual clinician level, your fact-check is the guardrail the system does not have built in. The AMA survey found 88% of physicians have concerns about AI even as they use it. This skill is how you act on that concern every time a recommendation comes across your screen.

Try It Now:

💭 Final Thoughts

What connects the Anthropic self-replication research, the Hugging Face hack, California's new laws, the AMA's governance framework, and the FDA's regulatory rethink? They are all responses to the same gap: AI systems that can pursue goals without meaningful constraints on how they pursue them. The self-replicating model spreads because nothing prevents it. The breach succeeded because access was not adequately bounded. The California laws exist because AI was making clinical decisions without requiring a licensed professional to retain authority over them. These stories look different on the surface. They are describing the same problem.

For clinicians, the relevant version of that problem is immediate and specific. When you enter a patient's health information into a general-purpose AI tool without a Business Associate Agreement in place, you have given that system access to data it was not built to protect. When you act on an AI recommendation without verifying it, you have removed the one oversight mechanism the system does not provide on its own: you. These are not hypothetical risks. They are the same pattern, at clinical scale, playing out in exam rooms and on call right now.

The good news is real. The regulatory environment is moving faster than many expected. California's laws take effect in January. The FDA is rethinking how it authorizes generative AI devices. The AMA is writing governance frameworks that tie accountability to evidence rather than to vendor claims. The UK's commitment to post-market surveillance will influence policy beyond its borders. The governance structures that will shape AI in medicine are being built during this window. The individual clinician's job in that moment is straightforward: hold the line until they arrive. Verify before you act. Protect what you share. Stay in the loop.

Let me know what you’d like to learn and I’ll include it in a future newsletter.

email me @ [email protected]