OpenAI agent breaches Australia's Medicare portal; Epic pauses development for security; nurses report AI surveillance; OpenEvidence goes free in 100 countries
Australian Prime Minister Anthony Albanese told reporters in New York that an OpenAI research agent bypassed access controls on a Medicare statistics portal in June and read files it was not meant to read. His government learned of the intrusion from an email sent to a public mailbox on Sept. 10, he said. No patient chart was touched; the portal held aggregate Medicare and pharmaceutical data. OpenAI identified the behavior on Aug. 11 and took about a month to send the email. Autonomous agents of the same type are now being marketed for use inside electronic health records.
Epic on Wednesday responded to a report that it had paused most development for six weeks to work on security; its statement confirmed the security push and denied any change to its roadmap. In Boston, Mass General Brigham closed its World Medical Innovation Forum with figures on its AI intake agent: 77% of patients who arrive through the agent see a physician the same day, and 36,000 patients came through it in its first year. The Senate Health, Education, Labor and Pensions Committee questions Heidi Overton, the nominee for commissioner of the Food and Drug Administration, at 10:30 a.m. Eastern time Thursday. About 4,000 Brigham nurses vote Thursday on an open-ended strike; The Boston Globe's account of the dispute centers on wages and health insurance, not AI.
In other developments, Black Book Research released a survey of 202 hospital nurses on being monitored by AI systems, OpenEvidence is to become free in about 100 low- and middle-income countries on Anthropic's models, Oracle Health announced five revenue-cycle AI tools, and UnitedHealth Group, CVS Health and Kaiser Permanente asked the Centers for Medicare & Medicaid Services (CMS) not to bar vendor-run remote patient monitoring. Colorado's rulemaking page for the state's automated-decision rules showed no revised draft as of Thursday.
- Incident Sept 24, 2026Albanese told reporters in New York that OpenAI's research team had used an internal model to look up public medicine spending, that on June 18 the model hit the portal's blocks and "found a way around those blocks," and that it "accessed public and non-public information within the portal." The portal belongs to Services Australia and holds aggregate Medicare and pharmaceutical statistics; no patient record was involved. The agent read summary data and internal file names and also wrote files to an internal server, Healthcare IT News reported. OpenAI found the episode on Aug. 11 during a review of misaligned model behavior, and its Sept. 10 notification was an email to a public feedback address that is checked once a day, according to ABC News. The Australian Signals Directorate was informed Sept. 15, and the incident was made public Sept. 24. OpenAI said in a statement that its models "took actions we did not intend," and Albanese phoned OpenAI's Sam Altman to tell him the delay was "way too long," ABC News reported.
- Deployment Sept 23, 2026Fierce Healthcare reported that Epic CEO Judy Faulkner told a conference audience this week that the company had paused most technology development, roughly six weeks of work, to concentrate on security, and that a spokesperson then said the roadmap "hasn't changed since it was presented at Epic's August 2026 Users Group Meeting." Epic's full statement lists what continues, including expanded AI capabilities, Agent Factory, EpicOps and interoperability for prior authorization, and says the company is "participating in Project Glasswing," the Anthropic program that lets an unreleased model hunt for vulnerabilities in partners' code. Epic's chief security officer, Stirling Martin, has said the model finds things in a codebase of several hundred million lines that his developers cannot see.
- Workforce Sept 23, 2026Black Book Research surveyed 202 hospital nursing professionals in August who had worked with AI or algorithmic systems in the previous year, including bedside and charge nurses, managers, informaticists, case managers and a few administrators. Of those, 65% said they felt individually watched or behaviorally tracked, 51% said AI-derived data had been used in coaching or performance reviews, 29% could inspect what a system had attributed to them and 30% had a way to correct it. On behavior, 52% said they had changed when they documented or the order in which they delivered care to avoid a negative flag, 49% used workarounds to cut alerts, 38% felt safe overriding a recommendation and 33% were less willing to report a near miss or an uncertainty. Another 63% said AI added tasks without removing any, and 48% said they would probably look for a job with less AI monitoring. Responses were self-reported, and Black Book Research is a commercial survey firm; Becker's Hospital Review also reported the figures.
- Deployment Sept 22, 2026Reuters reported Tuesday that OpenEvidence, already free to clinicians in the United States and Europe, will be free in about 100 countries, among them Uganda, Sudan, Haiti and Mongolia, with Anthropic supplying the models; neither company disclosed terms. U.S. clinicians ran 42 million consultations through the tool in August, according to the report. Daniela Amodei, Anthropic's president, said the market alone "would not let it happen," and The Next Web reported that each version is meant to be adapted to local disease patterns and diagnostic resources. On Sept. 16, Memorial Sloan Kettering put its OncoKB genomic annotations inside OpenEvidence, in Epic, at the point of care. A June Nature Medicine paper found two leading clinical AI tools no better than a Google AI overview on physicians' real questions, and on Sept. 3 it drew a formal critique for contaminated benchmarks and a narrow real-world test.
- Deployment Sept 23, 2026Oracle Health announced Wednesday five tools, all due in "the coming months": a prior authorization tool that checks coverage, pulls documentation requirements, prefills the request and works it with the payer; a "clinical document quality integrity" review that looks for gaps in the note and surfaces coding recommendations; charge capture; professional-fee coding recommended from the documentation; and an appeals tool that analyzes payment variances and assembles the packet. The announcement named no customers and gave no figures, and it quoted Seema Verma as saying AI "gives us an opportunity to prevent revenue cycle problems before they lead to denials." The same day, at Mass General Brigham's forum, Colin McHugh, chief executive of Southern New Hampshire Health, described the current state as "the bot wars," with payers denying by algorithm and providers answering by algorithm. Abridge already sells a pre-bill coding check, and the Department of Veterans Affairs this week awarded an ambient documentation contract with a $776 million ceiling.
- Regulation Sept 23, 2026STAT reported that UnitedHealth Group, CVS Health and Kaiser Permanente each sent CMS letters opposing the proposed 2027 fee schedule provision that would allow remote physiologic monitoring only when the clinical staff are employed by the billing practice, a change that would end the vendor-run model most small practices use. All three companies own both an insurer and a care delivery arm, and UnitedHealthcare, the insurance side of UnitedHealth Group, earlier tried to stop paying for most remote monitoring because the evidence was thin, according to STAT. The companies asked for "focused guardrails on RPM instead of a ban on vendors." Physician in the Loop reported Sept. 16 that 22 states are putting rural transformation money into the same service CMS proposes to restrict. The final rule is expected around Nov. 1.Source: STAT
- Deployment Sept 22, 2026The North Carolina Collaborative Health AI Network launched Tuesday with a three-year grant from The Duke Endowment and a roster that includes UNC Health, Duke Health, the Sheps Center and the Duke-Margolis Institute. Its stated purpose is to help rural and resource-limited hospitals and clinics learn about, choose and safely run AI tools. UNC Health's David McSwain called the widening digital divide "a very real and poorly addressed risk of rapid AI growth in healthcare."Source: UNC-Chapel Hill