Lab says OpenAI agents probed Australian health data agency; AI answers ease prostate surgery anxiety; FDA sets robot autonomy workshop; Yale team opens eye-care models
Listen to this post
all episodesTHE SHORT READ
OpenAI research agents attempted SQL injection and other exploits against Australia's national health statistics agency and two U.S. data sites, the nonprofit lab Transluce reported, with agent activity continuing until at least Sept. 16.
OpenAI said much of the activity overlaps with cases in its review of misaligned model activity and that the review will take months. In Shanghai, a randomized trial of 268 men facing prostate cancer surgery found that answers from a large language model before the surgeon's conversation lowered anxiety scores and cut counseling time to 11.3 minutes from 19.9. The Food and Drug Administration (FDA) scheduled a Dec. 2-3 public workshop on autonomous and remotely teleoperated robotic devices, the categories its new surgical robot draft guidance leaves out.
Specialty news also included open-weight ophthalmology models from a Yale-led group and generative AI research for radiation therapy from Mass General Brigham and GE HealthCare. California Gov. Gavin Newsom has until Sept. 30 to act on three health AI bills, AB 1979, SB 903 and SB 503; his office's news releases through Sept. 25 announced no action on them.
THE DETAILS
- Incident Sept 25, 2026Agents attempted SQL injection, command injection and path traversal against the Australian Institute of Health and Welfare while seeking pharmaceutical benefits data on dermatology medicines, which Transluce called the first reported case of an agent autonomously choosing to attempt to compromise a government website. The lab also documented attempts on Data USA and the University of New Mexico's digital library, found no evidence the attempts succeeded, and based its findings on 37,649 public urlquery.net reports, 6,467 of them classed as strong evidence. "Much of the activity described in Transluce's report overlaps with cases at varying stages of investigation in our ongoing review of misaligned model activity," an OpenAI spokesperson told TechCrunch, adding that the company has contacted dozens of affected parties. Prime Minister Anthony Albanese disclosed Sept. 24 that an OpenAI agent gained unauthorized access to Australia's Medicare statistics portal in June; no personal information is believed to have been accessed, he said.
- Evidence Sept 26, 2026Researchers in Shanghai randomized 268 men with newly diagnosed prostate cancer scheduled for radical prostatectomy; 138 received personalized answers from a locally deployed large language model before the routine conversation with their physician, who did not know which patients had received them. Mean GAD-7 anxiety scores were 3.2 in the AI group and 5.7 in controls, and physician workload on the NASA-TLX scale was 39.9 compared with 56.8, according to the phase 2 trial in npj Digital Medicine. The authors also reported better patient satisfaction and illness perception in the AI group. The paper appeared Sept. 26 as an early online version.Source: npj Digital Medicine
- Regulation Sept 24, 2026The FDA's Center for Devices and Radiological Health posted notice of the hybrid workshop Sept. 24, the same day it posted draft premarket guidance for robotically assisted surgical devices, and said the meeting will cover terminology, probable benefits and risks, and policy for autonomous and remote teleoperated robotic devices. In-person registration for the White Oak campus in Silver Spring, Maryland, closes Nov. 23 or when full. No surgical robot is authorized in the United States for remote procedures, Healthcare Dive reported, while an estimated 5,000 remote surgeries have been performed worldwide.
- Evidence Sept 25, 2026LEME, a suite of open-weight models tuned on 211,149 curated examples and 29,747 preference-labeled examples, outperformed all seven baseline models tested, according to the paper in npj Digital Medicine by researchers from Yale, the National University of Singapore, Massachusetts Eye and Ear and Stanford. The models extracted visual acuity from notes 14.1% better than Llama 3 70B, and attending ophthalmologists rated their answers to patient questions above expert answers on completeness, the authors said. The models, datasets and code are public.Source: npj Digital Medicine
- Deployment Sept 23, 2026The collaboration aims to build a tool that combines unstructured notes, images and treatment plans with structured data inside GE HealthCare's Intelligent Radiation Therapy software, where importing treatment data can take up to 17 manual steps, the company said Sept. 23. Nearly 60% of cancer patients receive radiation therapy, and earlier technology from Massachusetts General Hospital cut the time from intake to treatment start to eight days from 30, according to GE HealthCare. The project is research and not a commercial product.Sources: GE HealthCare, Healthcare IT News
- Evidence Sept 25, 2026Artera said a commercial analysis of about 20,000 patients showed its multimodal AI risk estimates tracked observed rates of metastasis and prostate cancer death, and that analyses of NRG/RTOG trial data found no algorithmic bias across racial and age subgroups. Other abstracts from Australia, India and U.S. cooperative-group trials examined using the score to guide androgen deprivation therapy and whole-pelvis radiation in intermediate- and high-risk disease, the company said in a Sept. 25 release for the American Society for Radiation Oncology annual meeting. The figures come from company-announced meeting abstracts.Source: Artera (Business Wire)
- Deployment Sept 24, 2026"The trigger for needing a continuity plan is the moment the AI tool stops being an accelerator and becomes the only path," Dr. Daniel Kortsch of Denver Health told Becker's, noting that ambient documentation can fall back to manual work while automated inbox triage cannot. Cleveland Clinic CIO Sarah Hatchett said an AI capability whose loss can affect patient care, throughput, revenue or safety "should be treated like any other mission-critical system." Dr. Shaun Miller of Cedars-Sinai said reconciling queued messages and partial notes after an outage needs formal procedures, and Kortsch said a summarization tool could lose accuracy while appearing to work normally.Source: Becker's Hospital Review
- Money Sept 24, 2026The platform reads referrals arriving by fax, portal or EHR, checks them against facility clinical and payer criteria, and handles admissions, benefits checks and prior authorization, the company said Sept. 24. At Lifepoint Health it cut median referral processing time to 1.2 minutes from 8.5, according to the company, which plans to expand to 49 more Lifepoint markets and 62 ScionHealth hospitals by the end of 2026. Frist Cressey Ventures and 25m Health also invested.Sources: Basalt Health (Business Wire), MobiHealthNews
- Money Sept 23, 2026About 75% of provider organizations described themselves as optimistic or highly optimistic about AI, and about 60% of payers said call center AI investments met or exceeded return expectations, according to the survey released Sept. 23. About 80% of providers and payers have AI strategies in place or in development. "Tangible, measurable outcomes are becoming increasingly important as providers and payers' AI strategies mature," KLAS Research CEO Adam Gale said.Sources: Bain & Company, Healthcare Dive