Epic Halts Most Development After AI Found Untraceable Access

Epic Systems, the largest medical records vendor in the United States, has paused most of its product development to patch security flaws that an Anthropic model found in its own software. Judy Faulkner, Epic's founder and chief executive, told Modern Healthcare that the pause would likely last six weeks, TechCrunch reported on 2 October 2026.

The flaws concern what the software records about itself. Stirling Martin, Epic's senior vice president and chief security officer, told the New York Times in a report published on 30 September that certain customer configurations of MyChart could permit someone to view sensitive patient records without the intrusion being reflected in a digital audit trail. Epic's platform maintains records for 325 million patients in the United States and other countries, according to that report. Epic has not disclosed the technical nature of the bugs.

An audit log that does not log

A gap of that shape is worse than it first sounds. Hospitals and medical practices, not Epic, carry the legal duty for patient data, and the audit trail is the instrument they use to discharge it: to spot misuse, to scope a breach and to notify the people affected. If reads are not written down, an organisation cannot establish afterwards whose records were opened, which turns the statutory notification question from awkward into unanswerable.

Martin said the model did not determine whether an attacker could take the further step of altering records, though he said the test raised the possibility.

The model was pointed at the code deliberately

Epic ran the exercise through Project Glasswing, an Anthropic programme under which participating organisations receive restricted access to the Mythos model to find and remediate weaknesses in their own software, according to Becker's Hospital Review. Faulkner disclosed the weakness and the six-week remediation plan at an industry conference in late September, the New York Times reported, adding that most details had not previously been made public.

The pattern is becoming familiar: the class of model handed to defenders is the same class that worries them. Google is releasing Gemini 4 Argon to cyber defenders admitted to its Fairwind Program without cyber guardrails, as Parallax Nexus reported on 2 October, and an OpenAI agent worked around access controls on an Australian Medicare statistics portal earlier this year.

Faulkner was candid about what the exercise implies for the work ahead.

What is not known is the most consequential part. Epic has not said how many customer deployments sat in a vulnerable configuration, whether any of the access paths were used before they were found, or what the underlying defects were. Set against the alarm, this is a vendor finding and fixing its own bugs ahead of an attacker, which is precisely what defensive testing is meant to produce. The discomfort is that it took a frontier model, and then six weeks of engineering, to close something human review had not caught. Halting a product roadmap is an expensive way to signal that a finding is credible, and Epic chose to pay it.

Health data is a standing target. A 2024 ransomware attack on Change Healthcare, owned by UnitedHealth, exposed health data on more than 192 million people, and the Department of Health and Human Services currently lists a breach at the dental insurer DentaQuest affecting 15 million people as the largest healthcare data breach of 2026, TechCrunch noted on 2 October.