1300 166 138 Tell us what you need
News / Security

Anthropic’s Mythos found a flaw in 320m health records

By Monika Tantau • Published 1 month ago
Two doctors review patient information on a tablet, the kind of medical record access exposed by the MyChart flaw Mythos found

Key takeaways

• Anthropic's Claude Mythos found flaws in some setups of Epic's MyChart patient portal
• Outsiders could view patient records without the access showing up in the logs
• Epic paused most product development for about six weeks to fix its software
• MyChart holds records for more than 320 million patients across the United States

Most Read

An AI model built to find security holes has found one in the software that holds the medical records of more than 320 million Americans. Anthropic’s Claude Mythos discovered that some setups of MyChart, the patient portal made by Epic Systems, could let outsiders view patient records without the access being recorded in the logs.

Epic’s response shows how seriously it took the finding. Founder and chief executive Judy Faulkner told Modern Healthcare that the company had paused most of its product development for about six weeks to focus on safeguarding its products.

What Mythos found inside MyChart

Epic’s chief security officer, Stirling Martin, told The New York Times that some customer configurations of MyChart could allow outsiders to reach patient records without any intrusion appearing in the software’s logs. An attacker who found the gap would leave no trace for hospital security teams to follow.

Martin said the testing also raised the question of whether records could be changed, not just read, though Mythos did not confirm that. There is no public evidence that anyone exploited the flaw before it was found, and the fixes have to be applied by the hospitals and clinics that run their own MyChart setups.

Why Epic was testing with Mythos

Epic had access through Project Glasswing, the programme Anthropic launched in April to let operators of critical software use Mythos on their own code before attackers find the same flaws. In June it expanded to about 150 more organisations in more than 15 countries, adding sectors such as power, water and healthcare.

Anthropic says partners have found more than 10,000 high or critical severity flaws with Mythos so far. It also expects many other AI companies to have models of this class within six to 12 months, which is exactly why defenders are racing to use it first.

A software engineer studies code on two monitors, as hospitals using Epic race to apply security patches

Patch, patch, patch

Martin’s advice to the hospitals running MyChart was blunt. He told the Times they need to get ready to patch, patch, patch, and that as soon as they think they are patching fast enough, they need to patch faster.

That is the new reality AI creates for defenders. When a model can find flaws that slipped past earlier reviews, the slow part becomes deploying the fix. Last week a Dutch security group showed the other side of that race, when an AI agent broke into its systems through two flaws nobody had reported.

Why Australian businesses should care

Epic is not only an American system. In Melbourne, the Royal Melbourne Hospital, the Royal Women’s Hospital, the Peter MacCallum Cancer Centre and the Royal Children’s Hospital share an electronic medical record built on Epic. Nothing published so far says any Australian setup is affected, but it is a reminder of how much patient data sits on a handful of platforms.

For clinics and allied health practices, the lesson is about logs and suppliers. Ask your software vendors how they test for flaws and how fast patches reach you, and check that access to records is logged. Our guide to AI tools for healthcare practices covers what to ask before adopting new systems.

AI is now the security auditor

The Epic case is one of the first public examples of an AI model finding a serious flaw in software that millions of people rely on. It will not be the last, and the same capability that helps defenders today will be in attackers’ hands soon after.

For organisations holding sensitive records, the practical response is to treat security testing as continuous rather than annual, and to plan for faster patching. That is a theme that comes up in almost every AI consulting conversation we have about health and finance data.

Scroll to Top