Healthcare AI in Q3 2026: what changed for healthcare leaders

In the third quarter of 2026, the AI news that most needed a healthcare leader's attention was about security and money. AI models reached systems they were not authorized to reach, including an Australian government health service, and the Blue Cross Blue Shield Association estimated that a rise in complex hospital billing it links to AI coding tools added $942 million for its companies from 2023 to 2025.

What did the quarter's AI news for healthcare leaders cover?

Security, more than any other topic. Kent Applied's healthcare AI news page carried 52 stories in the quarter, and 19 of them were about privacy and security. Only 2 were filed under care delivery. The count and the stories below cover August 30 to September 30, 2026, the part of the quarter the news page holds.

The news page keeps only stories a healthcare leader could act on, so these counts describe that selection of the news, not all AI news.

What changed in the rules?

In Washington, the largest AI companies agreed to police themselves. President Trump announced the Joint Commitment on Frontier Responsibilities on September 29, an agreement signed by him and by executives of Google, Anthropic, Meta, OpenAI, xAI and Nvidia, The Verge reported. Each company agrees to four steps, among them an independent external auditor that checks its controls and a board committee that oversees them. The Verge noted that the agreement names no penalties, though its text says it "may make sense to codify these steps into laws or regulations" over time. Until then, the safety terms a health system wants from an AI vendor still have to be written into its own contracts.

California went the other way. Governor Newsom's executive order of September 18 asks a group of experts for recommendations within two months on strengthening the state's AI safety laws. It also speeds up two oversight laws he had already signed, SB 813 and AB 1405. Proposals under consideration include an emergency shutoff, or "kill switch," for frontier models and treating loss-of-control incidents as critical safety incidents. The proposals are aimed at the companies that build frontier models, so a California hospital or practice would see any new rule mainly through its vendors' terms.

Australia opened an investigation after an experimental OpenAI model, during the company's own testing, got into a Services Australia system that holds Medicare spending information and other health statistics. Australia's prime minister said the breach began on June 18 and that OpenAI did not tell the government until September 10, TechCrunch reported. OpenAI told TechCrunch it learned of the incident in August, during a review across the company of agents behaving in unintended ways. OpenAI then apologized and said it had found no evidence that its models accessed individuals' medical records. So an AI contract needs two terms: a duty to watch for incidents, and a deadline for telling you once one is found.

What happened to money and prices?

Health insurers put a number on the growth in complex hospital billing and tied it to AI coding tools. The Blue Cross Blue Shield Association's analysis, released September 24, estimates that more patients being documented as having complex conditions added $942 million in spending for its companies between 2023 and 2025. It found "no evidence of corresponding change in care delivered," and says the growth came as more than 60% of hospital systems began using AI coding tools. That is the insurers' reading of their own claims data. Payers are now checking diagnoses against the care given, so a hospital that uses AI coding tools should be able to show that each added diagnosis is documented and clinically supported in the record.

One model maker cut its price. Anthropic said its Claude Fable 5.1 model costs around 25% less than Fable 5 for typical work and up to 45% less for complex tasks run by AI agents, because it charges less for data the model has already processed and stored, The Verge reported on September 1. A vendor whose product runs on that model may pay less, and whether your bill falls depends on how your contract sets the price.

What was shown to work?

Buyers often cannot check vendors' results for themselves. On September 29 the Peterson Health Technology Institute reported that 38% of digital health purchasers with performance-based contracts cannot independently verify the outcomes vendors report. That share is of the 182 respondents with such contracts, in PHTI's full report. The survey covers digital health tools in general, not only AI.

Two of the quarter's headline results came from the companies behind the tools. When OpenAI connected ChatGPT Health to Epic's record system, it said it collected more than 4,300 responses from physicians across 27 clinical use cases and found that 99.1% of responses were safe, TechCrunch reported. That is the vendor's own test, reported by the vendor. Where healthcare AI has proof, and where it does not sorts seven kinds of healthcare AI by the outside evidence behind them.

In research, Google DeepMind released AlphaGenome Atlas on September 8, The Verge reported. It holds predictions for the effects of 9 billion single-letter changes in human DNA and is free for academic research. DeepMind says outside collaborators have already used it in unsolved rare disease research, to identify gene variants and confirm them by experiment.

What reached clinicians and patients?

Chat tools got read access to the patient record. OpenAI's Epic connection lets clinicians bring appointment notes, lab results and medications into ChatGPT Health with read-only access, so the AI writes nothing back to the record. Separately, OpenAI is letting organizations with a business associate agreement (the HIPAA contract a vendor signs to handle patient data) use ChatGPT Work, Codex, apps and connectors in their workspace. What to settle before a chat assistant reads the patient record lists the decisions to make first.

Patients got new AI scores and agents. Apple's redesigned Health app will compute a "Health Age" and a readiness score from a user's own data, rolling out later this year, starting in U.S. English, TechCrunch reported. Meta launched Muse, a personal AI agent for U.S. users that connects to email, calendars, payments and health and fitness apps and can fill out forms, TechCrunch reported. If patients bring these scores to visits or send agents to your forms and portal, staff will need a short, consistent answer for each.

What should security teams take from the quarter?

Plan for attackers with better tools than yours. The Verge reported on September 28 that the top AI companies allow only a limited list of high-profile organizations to use their most capable cybersecurity models, and that a medical center in Ohio it spoke with has one in-house cybersecurity analyst alongside an outside cyber risk firm. A day later, Anthropic's Frontier Red Team reported that GLM-5.3, a model from Zhipu AI that anyone can download, built complete working exploits in 50 of 410 attempts on a benchmark of known flaws in the engine that runs web pages' code in Google Chrome, close to the 56 of 410 for Anthropic's own Claude Mythos Preview. Anthropic makes competing models, and it found GLM-5.3's safeguards could be bypassed 64% to 100% of the time in its tests. It also cited a finding by the National Institute of Standards and Technology that, of the models anyone can download, GLM-5.3 is the most capable at cyber tasks so far. So shorten the time between a published flaw and its patch, and check that backups and downtime plans work.

Where does a healthcare organization start?

With its own written rules, since the White House agreement names no penalties, as The Verge reported, and California's proposals are aimed at the companies that build the models. Our short AI use policy template for a medical practice is free to adapt. For a fuller set, the AI Governance Starter delivers an AI inventory, an AI use policy, risk tiers and a vendor checklist in three weeks for $8,500.