Safety prompt cuts harmful AI choices in 19 of 20 models; Oracle breach tally nears 20 million; device coverage comments close Oct. 13; Mayo AI grades side effects
Listen to this post
read by AI voices ยท all episodesTHE SHORT READ
A short safety reminder cut potentially harmful clinical choices by 20 open-weight language models from 16.6% to 10.1% across more than 10 million responses, Mount Sinai researchers reported in Communications Medicine.
The models still picked a harmful option in about 1.18 million responses, and the reminder helped least on cases drawn from real hospital records, where harmful choices fell from 46.6% to 24.5%, the authors reported. They called the work a proof of concept that does not estimate how often harm would occur in deployed systems.
Elsewhere, Bloomberg, citing a report from the Texas attorney general, put the toll of Oracle Health's 2025 breach of legacy Cerner servers at nearly 20 million people, according to Fierce Healthcare and SecurityWeek. Comments on Medicare's proposed fast coverage pathway for breakthrough devices close Tuesday. The weekend was otherwise quiet, and the post runs shorter.
THE DETAILS
- FYIEvidence Oct 8, 2026Researchers at the Icahn School of Medicine at Mount Sinai gave 20 open-weight language models 601 clinical scenarios, 100 of them adapted from real hospital records, in which an instruction pushed toward an unsafe choice, such as skipping recommended follow-up blood tests or stopping antibiotics early; the study, published Sept. 26 in Communications Medicine, logged more than 10 million responses. Adding a brief reminder to verify or escalate unsafe requests lowered harmful choices in 19 of the 20 models, but the rate on cases from real records still stood at 24.5% with the reminder, the authors reported. "A reminder should be viewed as one safeguard, not a substitute for clinical oversight," Dr. Mahmud Omar, the first author, said in the health system's Oct. 8 release. The authors said the study does not show how often harm would occur in practice.
- FYIIncident Oct 9, 2026Cerner's entry on the Texas attorney general's breach portal, dated Oct. 2, lists 2,992,244 affected Texans, and news reports citing the Texas filing put the national total near 20 million, according to SecurityWeek and Fierce Healthcare. A sample notice filed in California says exposed data may include names, Social Security numbers, diagnoses, medicines, test results and images, according to SecurityWeek. Oracle has not disclosed a total and did not comment, Fierce Healthcare reported. A federal judge in Missouri ruled in June that a consolidated class action can proceed against Oracle and eight health systems, holding that providers cannot be "absolved of liability" for protecting patient data by contracting it out, according to Becker's. Correction (Oct. 14): This item and the post's summary have been corrected. An earlier version said in the item's title and the post's summary that state filings put the breach's toll at nearly 20 million people; Bloomberg reported that figure, citing a report from the Texas attorney general, according to Fierce Healthcare and SecurityWeek, and Cerner's entry on the Texas attorney general's breach portal lists 2,992,244 Texans. The Becker's Hospital Review report on the June ruling, which the item cites, has been added to the item's sources.
- ActRegulation Oct 9, 2026
Comments on CMS's RAPID coverage pathway notice, docket CMS-2026-2674, are due Oct. 13, according to the Federal Register.
Under the Centers for Medicare & Medicaid Services notice, published Aug. 11, CMS would post a proposed national coverage determination the day the Food and Drug Administration authorizes an eligible breakthrough device, with a final decision about 60 days later for Class II devices and 90 days later for Class III devices. Devices must enter at the pre-investigational device exemption stage and run a study measuring outcomes in Medicare beneficiaries; laboratory and other in vitro diagnostic tests are excluded, and the earlier TCET pathway is paused for new candidates. The device industry group AdvaMed asked CMS on Oct. 8 to widen eligibility, admit diagnostics and set firm timelines. "Medicare beneficiaries shouldn't have to wait for FDA-authorized breakthrough technologies," said Scott Whitaker, its chief executive. The notice does not mention software or artificial intelligence. - FYIEvidence Oct 10, 2026Mayo Clinic radiation oncologists combined speech recognition with language models to grade 10 urinary and bowel toxicities on the National Cancer Institute's standard adverse-event scale from recorded conversations with 31 men receiving radiotherapy for prostate cancer, across 39 visits, according to the study published Oct. 10 in npj Digital Medicine. The tool matched the expert reference grade 82.6% of the time, with a weighted kappa of 0.717, against 0.681 agreement among the expert raters themselves, the authors reported. Most assessments were grade 0, which inflates raw accuracy; F1 scores for grades 1 and 2 were 0.58 and 0.57, and 32% of positive flags were false. The authors described accuracy at grade 2 or higher as unproven.Source: npj Digital Medicine
- FYIMoney Oct 8, 2026MMI announced the Series D round, led by BioStar Capital, on Oct. 8 and said it would spend the money on wider U.S. and international sales, energy instruments, digital surgery features and clinical studies. The Symani system filters hand tremor and scales a surgeon's movements to miniature instruments for open repair of small vessels and nerves, according to MedTech Dive; FDA granted it De Novo authorization in 2024. The company is also testing the robot in REMIND, a study of reconnecting lymphatic drainage in the neck in patients with neurodegenerative disease, which FDA allowed to expand to 15 patients this summer, according to MedTech Dive. No outcome data from that study have been published.Sources: MMI, MedTech Dive