AI Security Model Uncovers Flaws in Epic Patient Records

Epic Systems, the company behind the MyChart patient portal, has paused most of its product development for roughly six weeks after an Anthropic cybersecurity model surfaced flaws that could expose patient records. It is one of the clearest cases yet of AI doing security work at a scale human reviewers cannot match, and of the disruption that follows.

What happened

Epic founder and chief executive Judy Faulkner told Modern Healthcare that most new development was halted while the company works on safeguarding its products. The flaws surfaced after Epic pointed Mythos, Anthropic's frontier cybersecurity model, at its own software.

Chief security officer Stirling Martin told The New York Times that some customer configurations of MyChart could let outsiders read patient records without writing an entry to the software's access logs. The model could not confirm whether records could also be altered undetected, but Martin judged the risk serious enough to remediate.

MyChart carries more than 320 million patient records across US hospitals and clinics. Epic does not hold customer medical data itself, so a configuration flaw can be replicated across many provider deployments at once. Epic participates in Project Glasswing, Anthropic's programme that gives roughly 50 partners early access to Mythos for defensive work.

Why it matters

Volume is the story. In the first month of Project Glasswing, partners used Mythos to find more than ten thousand high- or critical-severity vulnerabilities, according to Anthropic. Cloudflare alone found 2,000 bugs across critical-path systems, 400 of them high or critical severity. Anthropic's own scans of over 1,000 open-source projects turned up an estimated 6,202 high- or critical-severity flaws, with 90.6 percent of the 1,752 externally reviewed findings confirmed as real.

Patch throughput has not kept pace. Anthropic reported in its Mythos capabilities assessment that fewer than one percent of the vulnerabilities the model surfaced had been fully patched by their maintainers. Some open-source maintainers have asked the company to slow down disclosures so they can absorb the queue.

The builder takeaway

• Treat AI-driven security review as an agent workload. An isolated container, a source tree, a prompt, and an agent that hypothesises, runs the software, and writes a proof of concept is now a realistic pipeline you can assemble.
• Automate triage, not just scanning. Thousands of findings a month break any manual queue. Deduplicate, score severity, route to owners, and open tickets without a human in the loop.
• Assume the same capability on the attacker side. If a model can find a decades-old flaw in hours, your patch window is shrinking faster than your release cycle.
• Treat audit logging as a security control. The Epic issue was partly about access that left no trace, which is the failure mode that makes everything else undetectable.