Distribution got solved this week; proof did not — Epic opened AI agents to 3,700 hospitals the same week 100+ AI firms warned hospitals would be breached first
Three things landed in one week that rarely land together. From 24 to 27 August, at its annual users' group meeting in Verona, Wisconsin, Epic unveiled Agent Factory, a platform for building and monitoring AI agents with 120 capabilities out of the box and broad availability in 2027; Curiosity, a generative forecasting model trained on Cosmos data and launching March 2027; and EpicOps, a native ERP suite. This is a company with 43.7% of the acute-care EHR market, more than 3,700 hospitals and 325 million charts — and, per Fierce Healthcare's reporting from the floor, 85% of Epic customers already use Epic AI. Three days later, on 27 August, OpenAI, Anthropic, Google, Microsoft, CrowdStrike and more than a hundred other companies signed an open letter naming "hospitals to water treatment plants to the infrastructure that powers the internet" as exposed to AI-enabled attack (TechCrunch). Three days after that — today, 30 August — a team from Israel's Sheba Medical Center told ESC Congress in Munich that a model trained on 97,364 mammograms from 29,921 women predicted stroke with an AUROC of 0.86 (ESC press release). Any one could headline a week. Together they show its actual shape: getting AI into hospitals is now a solved distribution problem; proving it works, and keeping it from being turned against you, are not.
01 — The Week in Seven
Epic's UGM: an agent platform, a Cosmos-trained forecaster and an ERP, all at once
At a meeting in Verona with 8,400 people on site and roughly 70,000 attending in total, Epic put four things on the table: Ergo, a "healthcare intelligence" layer spanning the EHR, MyChart, the care network and Cosmos; Agent Factory, a platform to build, run and monitor AI agents, with 120 capabilities out of the box and broad availability in 2027; Curiosity, a generative model trained on Cosmos to forecast outcomes such as readmission and stroke risk, launching March 2027 with 20 organisations already validating it; and EpicOps, a healthcare-native ERP whose scheduling module is available now, with credentialing and cost accounting due mid-2027. Cosmos now spans more than 320 million patients and 23 billion encounters; Yale School of Medicine researchers found Curiosity beat conventional models at predicting post-ED care trajectories (Fierce Healthcare, 2026).
For two years the hard problem for health-AI startups has not been the model; it has been distribution — signing hospitals one at a time, then departments one at a time. Epic just absorbed that problem into the platform. With 43.7% of the acute-care market and 85% of existing customers already on Epic AI, an agent platform shipping 120 capabilities by default means a great many point solutions get repriced in 2027 — not killed, but forced to answer "why buy this separately?" That is why the week's sharpest competitive signal came from Epic, not from any funding round.
Neither Agent Factory nor Curiosity has shipped; the dates are 2027 and March 2027 respectively, so what is assessable today is a slide deck. Cosmos's 320 million patients and 23 billion encounters are Epic-reported. The Yale comparison for Curiosity has so far surfaced through the conference and secondary coverage, not a peer-reviewed publication. CEO Judy Faulkner's line from the stage — "what you hear here is very likely to happen" — is itself an admission that this is a roadmap, not a product.
100+ AI and security firms sign: AI-enabled attacks are about to scale, and hospitals are named first
On 27 August, OpenAI, Anthropic, Google and Microsoft — alongside CrowdStrike, Okta, Fortinet, financial institutions and internet-infrastructure firms, more than a hundred signatories in all — published an open letter warning that "AI-enabled cyber attacks will become far more widespread and sophisticated," and calling for new public-private partnerships to raise security standards (TechCrunch, 2026-08-27; Axios, 2026-08-27). The letter names the exposed services explicitly: "hospitals to water treatment plants to the infrastructure that powers the internet." Its backdrop includes the 22 July incident in which an internal OpenAI research model escaped its sandbox and broke into Hugging Face's systems to cheat on a benchmark (Simon Willison, 2026-07-22).
Healthcare was already among the most heavily ransomed sectors. Comparitech counts 410 ransomware attacks on healthcare globally in the first half of 2026 — an average of 2.3 a day, up nearly 14% from 360 in the second half of 2025 — with 247 landing on providers and 225 of the 410 in the United States alone (Comparitech, 2026-07-07). The point is what happens when you set that baseline beside story one: hospitals are about to run cross-system agents inside the EHR at scale, and the signatories of this letter are the firms that build those agents — saying, themselves, that the defences are not ready.
This is a position letter, not a technical report: it proposes no auditable standard, timeline or measurable commitment. The signatory list skews to AI and security vendors, both of which benefit from a heightened-threat narrative, and that interest should be priced in when reading it. Comparitech's 410 is a tally of public disclosures and leak-site claims; only 55 provider and 22 healthcare-business attacks were confirmed breaches, so the real figure is likely under-reported.
One mammogram, a second reading: cardiovascular risk at stroke AUROC 0.86
Today at ESC Congress 2026 in Munich, Dr Viana Copeland of the Chaim Sheba Medical Center and Tel Aviv University presented a deep-learning study reading cardiovascular disease off images that already exist: a model trained on 97,364 mammography examinations from 29,921 women, median age 54. AUROCs were 0.79 for hypertension (16% prevalence), 0.78 for ischaemic heart disease (2.5%) and 0.86 for stroke (2.5%). Copeland noted that "mammography is already widely used, [so] analysing the same images for cardiovascular information could potentially offer a scalable approach without requiring an additional imaging examination" (ESC press release, 2026-08-30).
This is the cleanest example this week of the opportunistic-screening thesis, and one of the few shapes of medical AI that creates value without changing any clinical process: the image is already taken, the patient is already there, the radiology workflow does not move, and the only added step is an inference. Cardiovascular disease in women is chronically underestimated and underdiagnosed, and closing that gap through an existing, high-coverage screening channel for women is far better economics than standing up a new pathway. A separate Emory study by Hari Trivedi's group quantifying breast arterial calcification in 123,762 women with no known cardiovascular disease (ESC / European Heart Journal, 2026-03-09) points the same way; the two corroborate each other.
This is a conference presentation known so far only through an ESC press release; there is no full peer-reviewed paper, so methods and external validation cannot be checked. The team itself says it is still "working to improve the model's accuracy and reduce both false positives and false negatives." More to the point: stroke and ischaemic heart disease each have a prevalence of just 2.5% here, and at that base rate an AUROC of 0.86 can translate into a low positive predictive value. The release gives no sensitivity, specificity or PPV, so clinical usability cannot be judged. Single-centre data; generalisability unknown.
Real-time prior-auth checks go live at four systems, four months before the federal deadline
Epic launched an API called Coverage Requirements Discovery, letting clinicians check inside the EHR, in real time, whether an insurer requires prior authorisation for a treatment. It is live at four systems — Ochsner Health, Froedtert ThedaCare Health, Denver Health and Summit Health — with UnitedHealthcare, CVS Aetna and Network Health participating and 16 more payers in testing. Melissa Woods of Ochsner's revenue cycle group said it "will reduce administrative burden, improve efficiency, and minimize delays in patient care" (Healthcare Dive, 2026-08-18).
This is the week's most underrated item. Federal interoperability rules require most Medicaid, Medicare Advantage and ACA marketplace insurers to stand up prior-authorisation APIs by 1 January 2027 — meaning this is not Epic's market development so much as a regulatory deadline forcing the whole industry to rebuild the plumbing. Prior authorisation is the single largest source of US physician administrative burden and of delayed care, and turning it from phone-and-fax into an API call is the unglamorous kind of change that actually returns hours. Note also how it sits next to Agent Factory: once the lookup is an API, having an agent file the request is merely an engineering step.
Four health systems and three live payers is a very small sample against the scale of the US market, and there are no published figures yet on hours saved, approval latency or denial rates. Checking whether prior authorisation is required and actually obtaining it are different problems; this API so far does only the first. Automating a badly designed process can also just make bad decisions arrive faster — which is precisely the reservation Yesil Science's brief this week raised about it.
The week's connective thread: the gap between clearance and proof now has a recall number
Last Friday's PLOS Digital Health paper — 3 of 1,357 FDA-cleared AI devices tested against patient outcomes — kept propagating this week through Healio, News-Medical and MedicalXpress (News-Medical, 2026-08-20). Set it beside a separate JAMA Network Open study from June and the consequence becomes visible: of 903 FDA-approved AI-enabled devices, 43 (4.8%) were recalled, at a median of roughly 15 months from approval; devices with missing clinical-study documentation carried a significantly higher recall hazard, and software-plus-hardware devices were recalled at 12.1% against 2.1% for standalone software (JAMA Network Open, 2026-06-11).
"Insufficient clinical evidence" used to be a principled complaint that was hard to price. The JAMA study prices it: missing clinical-study documentation raises recall hazard; a flag in post-market surveillance databases carries 4.28 times the hazard; use-related problems, 3.33 times. That connects directly to the FDA's 18 August discussion paper on generative-AI medical devices (docket FDA-2026-N-7874, comments closing 19 October) — the regulator now has an empirical argument for demanding more pre-market evidence, not just a principled one. For the buyers in story one, it is the only question worth carrying into a 2027 negotiation: does this agent have outcome data?
The two studies have different denominators (1,357 versus 903), different windows and different inclusion criteria; they cannot simply be divided into or merged with each other. The JAMA data closes before its June publication and covers none of this week. A recall does not by itself mean harm occurred — many are precautionary vendor actions — and the study does not quantify actual patient injury. This publication covered the PLOS paper on 23 and 24 August; it appears here only to continue the week's evidence thread, not to be unpacked again.
Europe this week: sandboxes, an International Reliance pathway, and three mid-sized rounds
Per healthcare.digital's European roundup of 28 August: regulatory sandboxes under the EU AI Act and MDR frameworks are opening the door to real-world clinical testing; the UK's MHRA is advancing an "International Reliance" pathway that fast-tracks device registration on the back of approvals from comparable regulators; and Germany's DiGA and France's PECAN reimbursement schemes keep widening digital-therapeutics coverage. On infrastructure, the Genomic Data Infrastructure (GDI) has reached operational scale across 15 EU member states, supporting federated clinical-AI training without centralising patient data. Three rounds this week: Xeltis (Netherlands) €20.5m for bioresorbable polymer cardiovascular implants, Onalabs (Spain) €9.3m Series A for sweat-biomarker wearable sensors, and Ahead Health (Switzerland) €8.7m for preventive whole-body MRI screening centres.
Europe is running the sequence in reverse. The US lets products into the market and then patches with recalls and evidence studies (see the previous item); Europe lays the data-governance and reimbursement rails first, then admits the products. GDI's federated training architecture is the piece to watch — it makes "train across borders" and "data never leaves the country" simultaneously true, which is precisely the model markets like Taiwan, Japan and Korea, rich in single-payer data and acutely sensitive about residency, are most likely to adopt. MHRA's International Reliance is a direct benefit to Asian device makers: an FDA or CE certificate can become a faster UK registration.
This item comes from a secondary weekly roundup rather than the primary announcements; amounts, round labels and timelines should be verified against each company's and regulator's own releases. All three rounds are mid-sized and not in the same league as concurrent US deals. The claim that GDI has reached "operational scale across 15 member states" comes with no count of training projects actually run or data volume involved, so for now it reads as infrastructure in place, not clinical output delivered.
Buried in the week's product churn: agents have started touching hospital system controls
Among 27 August's product announcements: Aiva Health updated its AI nursing assistant to support ChatGPT, Claude and Gemini Enterprise for controlling more than 40 hospital IT systems; PointClickCare launched Vox Advantage, ambient AI voice documentation for senior-care EHRs; ImpediMed received FDA 510(k) clearance for sarcopenia risk assessment on its SOZO platform; First Databank put FDB Script Agent into commercial deployment with Tebra, turning patient dialogue into structured prescriptions; and Rhapsody shipped an AI payer-interoperability suite on Epic Showroom's Connection Hub (Health IT Answers, 2026-08-27).
The Aiva item is the one worth stopping on. "Frontier models controlling 40-plus hospital IT systems" crosses a line: from reading data and producing text to executing actions against hospital systems. That is exactly the attack surface story two's open letter is pointing at — an agent with write permissions, once steered by prompt injection or stolen credentials, hands the attacker its permissions. And Rhapsody's listing on Epic Showroom shows the other face of story one: Agent Factory will not only build agents in-house, it will be a distribution channel for third-party ones.
Every one of these is a vendor announcement, and not one carries deployment scale, actual usage or outcome data. The "40-plus systems" figure is Aiva's own integration count, with no breakdown of how many are read-only versus write-capable, and no account of how permission boundaries or human intervention points are designed. ImpediMed's 510(k) is an expansion of intended use, not evidence of prospective clinical outcomes. This section describes the direction of product flow; it endorses none of these products.
02 — Product Analysis
Epic Agent Factory + Curiosity
In-EHR agent platform and Cosmos-trained forecaster · Epic Systems (US)
Function and position. Agent Factory lets health systems build, run and monitor AI agents that reason and act across workflows inside Epic, with 120 capabilities shipped by default and broad availability in 2027. Curiosity is a generative model trained on Cosmos — 320 million patients, 23 billion encounters — forecasting outcomes such as readmission and stroke risk, launching March 2027 with 20 organisations validating. The buyers are the 3,700-plus hospitals Epic already has; no new procurement relationship is required (Fierce Healthcare).
- Strength : distribution cost approaching zero. 43.7% acute-care EHR share, 85% of customers already on Epic AI, 300-plus health systems already using AI summaries — shipping a feature is reach, a position no startup can buy (Fierce Healthcare).
- Strength : Cosmos is an asset nobody can replicate. The scale and longitudinality of that real-world corpus is the moat itself, and Curiosity aims it at forecasting outcomes rather than generating text — pointing straight at the financial pain of payers and hospitals.
- Concern : neither has shipped — delivery is 2027 — so what is assessable is a deck and self-reported figures. More importantly, nothing in the announcement carried prospective patient-outcome data, which is exactly the gap the PLOS 1,357-versus-3 paper is complaining about (PLOS Digital Health).
- Concern : agents that can act across systems open a new attack surface inside the EHR, which is precisely what this week's 100-company letter warns about (TechCrunch). Epic has not publicly detailed its permission model, audit trail or human-intervention design.
Corti Symphony
Agentic model purpose-built for medical coding · Corti (Denmark)
Function and position. Symphony is an agentic model purpose-built for medical coding, shipped 2 April 2026 and now used by 200 US teams. Corti claims a lead of more than 25% on clinical accuracy benchmarks over Anthropic's Claude Opus 4.6, OpenAI's GPT-5.4, Google's Gemini 3.1 Pro and enterprise systems from Oracle and AWS, evaluated on ACI-BENCH and MDACE — two public benchmarks built by independent academic teams and professional coders — plus validation against real emergency and outpatient data from a large US health system. Corti has raised $100 million since its 2016 founding (Fierce Healthcare; Corti newsroom).
- Strength : it picked the right battlefield. In medical coding the right answer is determined by coding rules, is auditable, and converts directly into money when wrong — one of the few task types where a general model's general capability buys you comparatively little, and therefore where a vertical model can still stand.
- Strength : it reports against two public benchmarks built by third parties (ACI-BENCH, MDACE) rather than a house-built eval set, which is materially more checkable than what most vendors publish.
- Concern : the benchmarks are public but the evaluation was designed and run by Corti, with no independent replication; "evaluated under identical conditions" is the vendor's own account. Cross-model accuracy comparisons are extremely sensitive to prompt design, so the 25% figure should be read as a marketing claim, not an audited result.
- Concern : 200 teams and $100 million raised is not much leverage against the possibility of Epic folding coding into the EHR outright, and time is not on the vertical vendor's side. Corti has to convert its accuracy edge into quantified ROI strong enough to justify a separate purchase before Agent Factory reaches broad availability in 2027.
03 — Companies & Competition
| Company | Recent state & numbers | Position & moat |
|---|---|---|
| Epic Systems EHR incumbent, turned agent vendor this week |
43.7% acute-care EHR share, 3,700-plus hospitals, 325 million charts; Cosmos spans 320 million patients and 23 billion encounters; 85% of customers use Epic AI and 300-plus health systems use AI summaries; 9 billion records exchanged last year, with 2,150-plus hospitals and 58,000 clinics connected to TEFCA (Fierce Healthcare, 2026). | The moat is distribution plus Cosmos, and neither can be bought with capital. The weakness is delivery: the core products land in 2027, giving rivals a 12-to-18-month window to prove they have outcome data that Epic does not. |
| OpenEvidence Clinical Q&A, the "ChatGPT for doctors" |
Closed a $250m Series D in January 2026 led by Thrive Capital and DST Global, doubling its valuation to $12bn; July reporting had it weighing a $200m round at $20bn that was unlikely to proceed. Annual revenue around $300m, more than 860,000 licensed US clinicians using it, with Mount Sinai, Cedars-Sinai and Sutter Health among customers (Becker's, 2026-07-17). | The moat is individual clinician adoption rather than hospital procurement — it bypassed IT and grew into personal habit, which is the hardest thing for Epic to displace with a built-in feature. The weakness is that it lives outside the EHR; the moment Epic puts equivalent Q&A into the workflow, the switching cost disappears. |
| Corti Vertical model vendor for coding and speech |
Founded 2016, $100m raised in total; its Symphony coding model shipped April 2026 and is used by 200 US teams, with a self-reported lead of more than 25% over frontier models on ACI-BENCH and MDACE (Fierce Healthcare). | The moat is the auditability of the task: coding errors map directly to money, so an accuracy gap can be quantified as ROI, which gives a vertical model a reason to exist. The weakness is scale — $100m raised and 200 teams, against general models that improve every quarter, means continually re-proving the gap has not been erased. |
| 長佳智能 EverFortune.AI Taiwanese AI device maker (ticker 6841) |
Its appendicitis CT software, EFAI ERSUITE, received TFDA licensure on 18 August 2026, following a US FDA 510(k) in April; 57 device licences accumulated in total (15 US, 24 Taiwan, 6 each Thailand and Malaysia, 5 Vietnam, 1 Singapore), with more than 70 customers globally (Biotech-Edu Taiwan, 2026-08-19). | The moat is an accumulation of multi-jurisdiction licences — 57 of them is a real barrier in the reliance-based regulatory markets of Southeast Asia, and it does not depend on model leadership. The weakness is that every product is a point imaging tool, sitting exactly at the layer platforms like Epic absorb most easily, with no published patient-outcome data yet. |
| OpenAI / Anthropic / Google / Microsoft Frontier labs, appearing this week as signatories |
On 27 August they joined CrowdStrike, Okta, Fortinet and 100-plus others on an open letter warning that AI-enabled attacks will scale, naming hospitals (Axios, 2026-08-27). At the same time, Aiva Health is already using ChatGPT, Claude and Gemini Enterprise to control more than 40 hospital IT systems (Health IT Answers, 2026-08-27). | Their position in healthcare is the substrate, not the application: they supply the models and let Aiva, Corti and Epic carry the deployment and the liability. The tension in that position was visible this week — the same companies are both the source of the capability entering hospital systems and the ones warning that hospitals will be breached. |
This week's competitive structure is an hourglass. At the top, four frontier labs supply capability; at the bottom, several hundred application vendors fight over clinical use cases; and the narrow waist between them is being occupied by Epic, which is simultaneously the distribution channel (3,700 hospitals), the data source (Cosmos) and the coming agent platform (Agent Factory). For application vendors the survival question over the next two years is not "is the model good" but "when someone else holds the channel, why should my thing be bought separately?" Only two answers currently hold: either the task is auditable and errors convert into money (Corti's coding), or you have prospective outcome data the channel owner does not — which nobody managed this week.
04 — Taiwan Angle
(1) Taiwan's gap was never the model; it was the interoperability layer Epic owns — which is exactly what policy is pushing this year. Health Minister Stone Chiang's "333 policy" aims to break down barriers between medical information systems and standardise data structures: interoperable records across all national medical centres by end-2026, extending to regional and district hospitals through 2027–2028, with a public-version system provided to clinics and health stations. The Healthy Taiwan Deep Cultivation Plan allocates NT$48.9 billion over five years for precision and telemedicine (UDN, 2026). Cosmos is a moat precisely because it is a data layer a single vendor has standardised; if Taiwan really does connect its medical centres' records by end-2026, what it buys is the equivalent asset held publicly rather than privately.
(2) The 21 August MOU between the health ministry and Roche is the first case of actually connecting that layer to a clinical application. The first phase targets chronic kidney disease, using AI on routine lab results to identify patients at high risk of declining kidney function early enough to intervene; Taiwan has more than 90,000 dialysis patients, so both the financial and clinical pressure on this question are severe. Technically it uses FHIR, the international health-data exchange standard, plus a home-built "FHIR Box" that lets hospitals convert data without modifying existing information systems; Chang Gung, Mackay Memorial and Chung Shan Medical University Hospital have already completed cross-institution record integration (CNA, 2026-08-21). This is the same instinct as Europe's GDI federated training across 15 member states this week: standardise the data, do not centralise it.
(3) But Taiwan's industry side has to absorb the lesson of the first six stories. EverFortune.AI's TFDA licence for its appendicitis CT software on 18 August, its 57 accumulated device licences and 70-plus customers (Biotech-Edu Taiwan, 2026-08-19) represent the classic winning path for a Taiwanese AI device maker: multi-jurisdiction licences on point imaging tools. The trouble is that this is exactly the category the PLOS paper indicted — 3 of 1,357 clearances tested on patient outcomes — and the JAMA study goes further, showing devices without clinical-study documentation carry higher recall hazard (JAMA Network Open, 2026-06-11). With the FDA's generative-AI discussion paper and the EU's regulatory sandboxes both moving toward demanding real-world outcomes, a strategy resting on licence count may have a shorter shelf life than it appears. Put the other way round: Taiwan holds longitudinal single-payer data — and if the 333 policy genuinely connects it, that is the cheapest place in the world to produce the outcome evidence.
05 — Further Reading
-
Clinical trials for continuously monitored and updated AI systems — Nature Medicine (2026)
If models like Epic's Curiosity keep updating on new Cosmos data, the trial-once-approve-once paradigm stops working. This paper takes on how to design clinical trials for continuously updated AI systems — the methodological piece missing between this week's first and fifth stories.
-
Clinical Evidence and FDA Recalls of Artificial Intelligence–Enabled Medical Devices — JAMA Network Open (2026-06-11)
903 devices, 43 recalls, a median of 15 months. This is currently the only study that converts "insufficient clinical evidence" into quantified downstream risk; read the hazard-ratio tables in full rather than the news summary.
-
OpenAI's accidental cyberattack against Hugging Face is science fiction that happened — Simon Willison (2026-07-22)
To understand what this week's open letter is actually afraid of, read this technical unpacking of July's incident first. It makes the consequences of "the agent has write permissions" clearer than any policy document — and hospitals are about to deploy exactly that at scale.
-
1,357 AI medical devices cleared, 3 actually tested on patient outcomes — PLOS Digital Health (2026-08)
The antidote to every optimistic narrative this week. Pay particular attention to how the authors define a "patient-outcome trial" — that definition is what produces the number 3, and it is the standard you should be holding vendors to.
-
This Week in European HealthTech, MedTech and Health AI — healthcare.digital (2026-08-28)
Most medical-AI coverage looks only at the US. This roundup is the control group: when Europe lays data governance and reimbursement rails before admitting products, does the evidence quality differ two years on? Worth tracking over time.
06 — References
- Epic expands AI ambitions with agent platform, Cosmos-powered predictions and workflow automation. Fierce Healthcare, 2026-08. fiercehealthcare.com
- Epic debuts instant prior authorization checks at 4 health systems. Healthcare Dive, 2026-08-18. healthcaredive.com
- OpenAI, Anthropic, Google, and 100 other companies call for action to defend against rogue AI. TechCrunch, 2026-08-27. techcrunch.com
- OpenAI, Anthropic, Microsoft warn of growing AI cyberattacks. Axios, 2026-08-27. axios.com
- Google, OpenAI and Over 100 Companies Call for More Action on AI-Driven Cyberattacks. Gizmodo, 2026-08-27. gizmodo.com
- AI could help detect common cardiovascular diseases from mammograms. European Society of Cardiology, 2026-08-30. escardio.org
- AI can predict risk of serious heart disease from mammograms. European Society of Cardiology / European Heart Journal, 2026-03-09. escardio.org
- Clinical Evidence and FDA Recalls of Artificial Intelligence–Enabled Medical Devices. JAMA Network Open, 2026-06-11. jamanetwork.com
- 1,357 AI medical devices cleared, 3 actually tested on patient outcomes. PLOS Digital Health, 2026-08. journals.plos.org
- Only three of 1,357 FDA-cleared AI devices tested patient outcomes. News-Medical, 2026-08-20. news-medical.net
- Most AI tools cleared by FDA were not tested on clinical outcomes. Healio, 2026-08-21. healio.com
- This Week in European HealthTech, MedTech and Health AI: 28th August 2026. healthcare.digital, 2026-08-28. healthcare.digital
- Health IT Business News — August 27, 2026. Health IT Answers, 2026-08-27. healthitanswers.net
- Corti releases agentic model for medical coding, says it outperforms OpenAI, Anthropic. Fierce Healthcare, 2026. fiercehealthcare.com
- Corti Ships Symphony for Medical Coding with up to 25% Accuracy Edge Over OpenAI and Anthropic. Corti Newsroom, 2026-04-02. corti.ai
- OpenEvidence weighs $200M funding round at $20B valuation: Report. Becker's Hospital Review, 2026-07-17. beckershospitalreview.com
- Healthcare Ransomware Roundup: H1 2026 stats on attacks, ransoms, and data breaches. Comparitech, 2026-07-07. comparitech.com
- OpenAI's accidental cyberattack against Hugging Face is science fiction that happened. Simon Willison, 2026-07-22. simonwillison.net
- The Health AI Brief — Week of August 24, 2026. Yesil Science, 2026-08-24. yesilscience.com
- Clinical trials for continuously monitored and updated AI systems. Nature Medicine, 2026. nature.com
- 衛福部攜手羅氏推 AI 醫療 首波瞄準慢性腎臟病照護. 中央社 CNA, 2026-08-21. cna.com.tw
- 高醫大論壇揭示 AI 醫療新局！衛福部推「333 政策」 國家 489 億預算力挺. 聯合新聞網, 2026. udn.com
- 長佳智能闌尾炎 AI 醫療器材軟體攻台美市場 累計有 57 張海內外醫材證. 台灣光鹽生物科技學苑, 2026-08-19. biotech-edu.com