Downtime is a clinical event
When ransomware locked hospitals' computers, more of the Medicare patients already in their beds died. As care depends on more software, AI included, keeping that software safe must come before speed or savings.
Listen to this special topic
read by an AI voiceRansomware attacks on hospitals endanger the patients already inside them. Ransomware locks staff out of their computer systems and demands a ransom to restore access. Researchers who linked a database of hospital ransomware attacks to Medicare claims have now measured what follows: in a study published in February, they found that among patients already admitted when an attack began, the in-hospital death rate rose by more than a third.
The study is observational. Compared with matched hospitals in the same states that were not hit, the death rate, which averaged under 4 percent, rose by about 1.3 percentage points. A comparison with hospitals attacked later gave a similar result. The study covers only patients in traditional Medicare, and the deaths in question are deaths in the hospital. Such attacks are common: the American Hospital Association says the FBI reported 460 ransomware attacks on health care in 2025, more than on any other sector of critical infrastructure. When a hospital's systems go down, by attack or by failure, the result is a patient-safety event, and as care comes to depend on more software, artificial intelligence included, keeping that software safe and running must come before speed or savings. That calls for firm security minimums for hospitals and the companies that hold their records, public money for the hospitals that cannot afford them, and medical staffs ready to work without the screens.
The harm spreads beyond the hospital under attack. When ransomware struck a neighboring health system's hospitals for a month in 2021, academic emergency departments in San Diego County took in more ambulances and more patients, including more than twice as many with confirmed strokes; more than twice as many people left without being seen, and waits grew longer. The study compared the weeks before and during the attack in one community, and its authors suggest that such attacks should be treated as disasters, with coordinated planning and response.
Other outages show the same disruption, though not always the same toll. A retrospective analysis of the WannaCry attack of 2017 found that infected hospitals in the National Health Service admitted about 6% fewer patients a day; it found no difference in deaths in emergency departments, which its authors called a crude measure of harm. A failure can do similar damage without any attacker. In July 2024 a faulty software update from the security firm CrowdStrike took down network services at about a third of the hospitals whose systems researchers could scan, and more than a fifth of the services that went down supported direct patient care. The authors note that generalizable data on what such outages do to patients are lacking.
In London the harm has been counted. After a ransomware attack in June 2024 on Synnovis, which provides blood testing and other pathology services to the National Health Service, the health service's managers confirmed 170 incidents in which patients came to harm, one of them severe, the Health Service Journal reported. King's College Hospital later said that a long wait for a blood test result caused by the attack had contributed to a patient's death. In Vermont in 2020, an attack left the University of Vermont Medical Center able to give chemotherapy to fewer than half the patients a day it normally could, its president said.
Not every death blamed on an outage was caused by one. In 2020 a woman turned away from Düsseldorf's University Hospital, which hackers had crippled, died after being diverted to another hospital; a chief public prosecutor in Cologne investigated and concluded: "The delay was of no relevance to the final outcome." In Alabama, a mother sued the hospital where her daughter was born during a 2019 ransomware attack, alleging that a doctor could not properly monitor the delivery; the baby suffered severe brain injuries and died the next year, and the hospital denied wrongdoing but later settled the suit.
Artificial intelligence raises the stakes on both sides. The hospital association told the Senate that governments and criminal hackers are beginning to use it to find software flaws quickly, and Anthropic reported last year that an attacker had used its coding tool in a data-theft and extortion campaign against at least 17 organizations, including in health care. Defenders use it too: testing with an Anthropic model turned up flaws in Epic's software that could allow undetected access to patient records, The New York Times reported. The evidence on downtime comes from record systems and networks rather than from AI, but its logic carries over to any tool that care depends on. ECRI, an independent patient-safety nonprofit, ranks misuse of AI chatbots first and unpreparedness for a sudden loss of electronic systems second among this year's health technology hazards. Guidance from the Joint Commission and the Coalition for Health AI tells health care organizations to treat AI-related safety events "like patient safety events."
Some of the defenses are basic, and a bill in Congress would require them. The Health Care Cybersecurity and Resiliency Act, which passed the Senate by unanimous consent on Sept. 30, would direct the health department to require multifactor authentication, a second check beyond a password, and encryption of patient information, among other minimums. The rules would cover health care providers and health plans, the vendors that handle patient records for them and other private organizations in the health sector, and would take effect three years after enactment. Vendors matter because one can fail many hospitals at once: the attack on Change Healthcare, which the hospital association says touches one in three patient records, disrupted patient care for months. The House received the bill on Oct. 5 and has yet to take it up.
Providers have resisted security mandates before. Late last year a group of health care organizations urged the health department to withdraw the security rules it had proposed in January 2025, saying they would place "substantial new financial burdens on health care providers." Every dollar a small hospital spends on security, the argument runs, is a dollar it cannot spend on care, and the deaths rest on a single observational study. The cost is real, and one study is not the last word. But every study cited here found disruption, to care or to the technology behind it. The American study that counted deaths found more of them, and in London the harm was confirmed case by case. The minimums also match how two recent vendor breaches began. Change Healthcare's attackers got in with someone else's login through a portal that did not have multifactor authentication, its parent's chief executive told senators. Oracle has told customers that the intruder in its health-records breach used stolen customer credentials. That breach may have exposed the records of nearly 20 million people, a reported count Oracle has not confirmed. Where a small hospital cannot afford the minimums, the public should pay, which is what the bill's grants are for.
The House should pass the bill with its coverage of vendors intact. Its grants, which name no sum, need money behind them: the hospital association says nearly half of rural hospitals lost money in 2023. States need not wait, since the $50 billion Rural Health Transformation Program already lets them pay for software, hardware and technical help that enhances cybersecurity.
Hospitals can act sooner still, writing multifactor authentication, encryption and regular testing into vendor contracts as they renew them, AI vendors included. For physicians the test is simpler: no clinical tool, AI or otherwise, should go live until the medical staff knows how care continues when it fails. The Joint Commission already advises hospitals to be ready to run with life- and safety-critical technology offline for four weeks or longer, with a downtime committee that includes the medical staff, regular drills and, afterward, root cause analyses of the consequences for patient care. Medical staffs should hold their hospitals to that plan as they would to a plan for mass casualties, and review every delayed result and diverted patient as a patient-safety event, so that the harm is counted rather than guessed at.
The patients in the Medicare study were already in their beds when the screens went dark, and more of them died in the hospital than would otherwise have been expected. As more of their care runs through software, AI included, keeping that software safe and running is part of treating them, and it comes before speed or savings.
- Ransomware and deaths in Medicare claimsHannah Neprash, Claire McGlave and Sayeh Nikpay, "Hacked to Pieces? The Effects of Ransomware Attacks on Hospitals and Patients," American Economic Journal: Economic Policy, vol. 18, no. 1, pages 256-81, February 2026, using fee-for-service Medicare claims. The abstract reports that ransomware attacks decrease hospital volume by 17 to 24 percent during the initial attack week, with recovery occurring within 3 weeks, and among patients already admitted when an attack begins, in-hospital mortality increases by 34 to 38 percent. The supplemental appendix lists 148 attacked hospitals and 873 matched non-attacked hospitals in the same state and a different Hospital Referral Region; its difference-in-differences estimates for in-hospital mortality are 0.0127 (dependent variable mean 0.0374) with matched hospitals as controls and 0.0140 (mean 0.0364) with hospitals attacked at least five weeks later. A working-paper version, supported by the NIHCM Foundation and using 2016-2021 claims, estimated a smaller relative increase for patients already admitted (20.7%, P=0.07) and found no similar increase in 30-day mortality.
- Emergency departments next to an attacked systemDameff, Tully, Chan and colleagues (JAMA Network Open, published online May 8, 2023) conducted a before and after cohort study of 2 urban academic emergency departments, with 11% of San Diego County's total acute inpatient discharges, during a month-long ransomware attack on 4 adjacent hospitals of a separate organization that constitutes approximately 25% of regional inpatient discharges. The periods were April 3 to 30, 2021 (before), May 1 to 28, 2021 (attack and recovery) and May 29 to June 25, 2021 (after), with 19,857 visits in all. The abstract reports increases in the attack phase in daily census, emergency medical services arrivals (1,741 vs 2,354), admissions, patients leaving without being seen (158 vs 360), median waiting room time (21 vs 31 minutes), stroke code activations (59 vs 102) and confirmed strokes (22 vs 47). The authors wrote that such attacks should be considered a regional disaster.Source: JAMA Network Open
- The attack on a London pathology contractorSynnovis, a contractor that provides blood testing, transfusion and other pathology services to the UK's National Health Service, was targeted in June 2024 by hackers who use the name Qilin, Bloomberg reported, and medical facilities were forced to postpone more than 10,000 appointments and cancel more than 1,700 elective procedures. On June 18, 2025, the Health Service Journal reported that managers had confirmed nearly 600 incidents, of which 170 involved patient care suffering: one case of severe harm, which prompted a patient safety incident investigation at King's College Hospital NHS Foundation Trust, 14 classed as moderate and 155 as low harm. A week later the trust said the hack was a contributing factor in the death of a patient, and that a long wait for a blood test result due to the cyberattack was a contributing factor. Synnovis's chief executive, Mark Dollar, said the attack had been identified as one of the contributing factors.
- Other outages, measuredGhafur and colleagues (npj Digital Medicine, Oct. 2, 2019) analyzed Hospital Episode Statistics data on the 2017 WannaCry attack on the National Health Service, comparing infected and non-infected hospitals by difference-in-differences. Infected hospitals had a decrease of about 6% in total admissions per hospital per day, with 4% fewer emergency and 9% fewer elective admissions; no difference was noted in deaths in emergency departments, the study's only mortality measure, which the authors called a crude measure of patient harm. Tully and colleagues (JAMA Network Open, July 19, 2025; corrected Aug. 13, 2025) scanned the internet services of 2,232 hospitals affiliated with US organizations that use the Epic record and have publicly observable services. After the faulty CrowdStrike update of July 19, 2024, 759 (34.0%) stopped responding; of 1,098 network services with outages, 239 (21.8%) were direct patient-facing services and 631 (57.5%) could not be classified. The authors wrote that generalizable clinical outcome data associated with such downtimes are lacking.
- Cases in Vermont, Germany and AlabamaVermont: Vermont Public reported on Nov. 13, 2020, that Stephen Leffler, president of UVM Medical Center, then recovering from a cyberattack, said it could give chemotherapy to about 15 patients a day, against 40 to 60 normally. Germany: Fortune reported on Nov. 16, 2020, that a woman turned away from Düsseldorf's University Hospital, crippled by hackers, died on the drive to another hospital (its Sept. 18 report said her care was reportedly delayed by an hour); Cologne chief public prosecutor Markus Hartmann said the delay was of no relevance to the final outcome. Alabama: CBS News and the Associated Press reported on Oct. 1, 2021, that a mother sued Springhill Medical Center, alleging that during a ransomware attack, when her daughter was born on July 17, 2019, a doctor could not properly monitor the delivery; the girl died at 9 months, and the hospital denied wrongdoing. Bloomberg reported in June 2025 that the hospital later settled.
- AI, and what safety bodies adviseAnthropic reported on Aug. 27, 2025, that an actor used its Claude Code tool for large-scale theft and extortion of personal data, targeting at least 17 organizations, including in healthcare. Becker's Hospital Review reported on Oct. 1, 2026, citing The New York Times, that Epic is patching flaws that could allow undetected access to patient records after testing its software with Anthropic's Claude Mythos model. ECRI ranked misuse of AI chatbots first, and unpreparedness for a "digital darkness" event, or a sudden loss of access to electronic systems and patient information, second among its 2026 Top 10 Health Technology Hazards. Guidance from the Joint Commission and the Coalition for Health AI (2025) says organizations should treat AI-related safety events like patient safety events. The Joint Commission's Sentinel Event Alert 67 (Aug. 15, 2023) says organizations should be prepared to have life- and safety-critical technology offline for four weeks or longer; drills can occur annually or quarterly.
- The Senate bill and rural moneyThe Health Care Cybersecurity and Resiliency Act of 2026 (S. 3315) passed the Senate with an amendment by unanimous consent on Sept. 30, 2026; the House received it and held it at the desk on Oct. 5. Section 8 directs HHS to update its HIPAA security regulations to require non-governmental entities in the Healthcare and Public Health Sector and covered entities and business associates to adopt minimum risk-based cybersecurity practices, including multifactor authentication, encryption of protected health information and monitoring, including penetration testing, effective 36 months after enactment. Section 10 lets HHS award grants of up to 3 years to federally qualified health centers, Indian Health Service facilities, nonprofit hospitals, rural health clinics and partner nonprofits; the text names no amount. The hospital association says 48% of rural hospitals operated at a financial loss in 2023. The Rural Health Transformation Program provides $50 billion over fiscal years 2026 to 2030; approved uses include technology designed to enhance cybersecurity capability development.
- Breaches, logins and costsThe hospital association says that in 2025 the FBI reported 460 ransomware attacks on the healthcare sector, more than on any other critical infrastructure sector; the next, critical manufacturing, had 355. It says Change Healthcare, a UnitedHealth Group subsidiary that touches 1 in every 3 patient records, was hit by ransomware in February 2024, disrupting patient care and hospital finances for months. UnitedHealth Group's chief executive, Andrew Witty, told the Senate Finance Committee on May 1, 2024, that criminals used compromised credentials on Feb. 12 to access a Citrix portal that did not have multi-factor authentication. On Oct. 8, 2026, SecurityWeek reported that Bloomberg, citing a Texas attorney general report, put those affected by the Oracle Health breach at nearly 20 million; Oracle has not made a public statement on the number. In a letter dated Dec. 8, 2025, organizations representing clinicians, providers and others urged HHS to withdraw its proposed security rule, citing substantial new financial burdens.