#94. AI is finding flaws faster than healthcare can patch them

The most important health tech story this week is that Epic, the largest EHR vendor in the world, stopped most of its product development to fix security flaws in its product, and told customers to “get ready to patch, patch, patch.”
The same week, NHS England’s board held its cyber security risk at 25, the highest score it can give. Put those two things together and you can see the problem: AI is now finding vulnerabilities faster than health systems can patch them.
What happened
Epic used Anthropic’s Claude Mythos, through Project Glasswing, to simulate attacks on its own software (TechCrunch, Becker’s). It found configurations of MyChart that could let outsiders view patient records without leaving any entry in the audit log. Epic has paused most new development for about six weeks while it fixes them. Around 320 million patient records sit on Epic systems.
The missing audit trail is a detail that will worry CIOs and clinical safety officers. An audit log is how a hospital knows who looked at what record. A breach that leaves no trace in the audit log asks a wider set of questions.
This follows from two earlier WIRTW issues. In #73 I asked what models like Mythos would mean for healthcare. In #92 an OpenAI agent worked its way around the access controls on an Australian government health website. It is now clear that frontier models are very good at finding the gaps in systems that many had assumed were closed.
Why it lands here
The size of Epic’s footprint here is what turns a vendor patch into a local problem. Across the UK, Epic is the electronic patient record for a number of large NHS trusts and regional health systems. Its MyChart patient portal is used by hundreds of thousands of patients here, and in one region it already reaches around one in seven adults.
Even if UK setups turn out not to be affected, NHS England’s own board papers show why the patching pace matters here. Cyber risk sits at maximum, and Data Security and Protection Toolkit returns show widespread “standards not met”. If many organisations are not meeting today’s baseline, a faster run of critical patches will be hard to absorb.
Patching is not free. Every release to a clinical system needs regression testing and change control, done by the same small digital and cyber teams who are already stretched. When a vendor says “patch, patch, patch”, the limiting factor is not the fix. It is how quickly each organisation can safely deploy it.
Patch speed is now part of keeping records private
Patients share their records on the understanding that only the people caring for them will look. When an AI model can find a way into a patient portal that leaves no audit trail, how fast an organisation can close it becomes part of how it keeps that promise.
There is a safety dimension too, but it sits in the patching rather than the breach. A rushed patch to a clinical system can break something clinicians depend on, which is why DCB0160 expects a clinical safety review of changes.
There is an upside. Health organisations can use the same AI tools to probe their own systems for weaknesses, and find them before an attacker does. But finding the flaws is only half the job. Each one still has to be fixed, tested and rolled out, and the real question is whether organisations have the staff and processes to do that at the speed AI can now find problems.
Also worth reading
- Ireland’s AI and EHR direction. At the All-Ireland Clinical AI Congress, the HSE’s Damien McCallion set out four pillars for “AI for Care”: clinical care, operations, research and population health.
- Ambient voice at scale. Barts Health now has more than 2,000 clinicians with access to Oracle’s ambient AI. 1,350 have used it and 700 used it in September. That gap between access and active use is the most honest AVT number I have seen, and a useful reality check on headcount-based claims.
- Guernsey winds down its EPR programme after spending £22m of a £23.7m budget, moving away from “big bang” towards incremental delivery (HTN). A small-jurisdiction warning worth reading.
- RFK Jr says AI is “much better informed than any doctor”, and six US medical bodies push back (Fierce Healthcare). Our patients use the same consumer tools, so the accountability question arrives here regardless of US politics.
- When AI shapes patient expectations (Medical Independent). A medico-legal view on what happens when a patient’s chatbot disagrees with their clinician.