UnHack with Drex DeFord

October 12, 2026: Bill Russell and Drex DeFord. Sarah Richardson is in the air for Philadelphia. The week is one argument: the attack got cheap, and the controls around the tool did not keep up.

The number is $700. Drex described a dark-web package that weaponizes a fresh patch for a buyer who does not understand it. At the regional staff roundtables, the matching line was a meme they were no longer joking about. They patch in production, with a plan to roll back, because the test window is gone.

Citrix had shipped another emergency NetScaler build that Sunday for a SAML sign-in bug attackers were already using. Epic had paused most new product work for about six weeks after Anthropic's testing found MyChart audit log gaps. Some customer configurations could let an outsider view records with no audit trail. Judy Faulkner is the source of the pause, via the TechCrunch and Becker's reporting Bill read on the tape. Mayo had given an all-clear after OpenAI agents probed the site on October 2. No unauthorized access. The agents looked for config files and tried to create accounts.

The policy block is the contract. Drex's takeaway from three stories in 2 Minute Drill Extra: get it in the contract, and a general counsel who does every other kind of agreement may need a specialist for this one. FDA put AI device lifecycle on the fiscal 2027 agenda. The FTC is looking at OpenAI and Anthropic. California signed AB 1979 and SB 503, protecting clinician judgment and requiring bias mitigation. Newsom vetoed the bill that would have protected a clinician who overrides the tool. The physician is still the person who owns the click. Drex's rhyme: EHR alerts trained a generation to click through, and AI alerts are walking the same road.

Children's Healthcare of Atlanta gates every Epic AI feature on accuracy and has left some off. Bill sat in a room where a CIO turned all of them on because the partner is trusted. Abridge is re-signing a great majority on three-year and five-year terms. Drex does not like predicting next year from a three-year signature.

HCA nurses report Timpani, built with Palantir, is staffing unsafely across about 130 of 190 locations. One union unit filed a grievance. Bill is not surprised by a prompt written to optimize.

Blue Cross Blue Shield's estimate is about a billion dollars over two years for autonomous inpatient coding. Bill's counter is the work was already delivered and the note is finally complete. Drex agrees the cost goes up because the payer has to pay for work that used to fall out of the chart. The next move he expects is delay: deny, demand the evidence, wait.

The story is the addition. The electrical plan passed. The electrician's questions did not. Cognitive surrender is Bill trusting the tool on the parts he will not learn. Drex's version from a dinner question: a lot of people who "have an agent" still have a search box.

Key Points:

0:00 Patch it in production

3:19 Citrix shipped the NetScaler patch

6:49 SureTest

7:32 Epic paused the new products

8:51 Past the business associate agreement

11:59 California, and the physician still decides

12:48 Censinet

15:40 Cognitive surrender

18:33 One Epic AI tool at a time

20:54 130 of 190. Nurses pushed back

23:01 A billion dollars over two years

25:54 A lawyer's inbox, and an agent

LinkedIn: 229Project

Donate: Alex's Lemonade Stand: Foundation for Childhood Cancer

October 5, 2026: Venkat Kavarthapu, CEO of Symplr, joins Bill Russell and Drex DeFord. The argument is about where AI time goes. Bill cites a study that put individual productivity up about 20 percent, and asks whether a faster IT administrator sees another patient. Drex's sticking point is the missing link to the CFO: less spend, or more revenue.

Venkat's picture comes from a CFO at a large system. Ten to 15 percent of the beds sit unused, on lock, unable to be scheduled. He puts the cause upstream. Credentialing a physician takes 120 to 180 days. Shrink that, and the capacity can open. Bill's answer is the buying rule. If that is what AI does, buy more of it.

Bill then reads a September 14 piece on health systems cutting jobs, including IT outsourcing at Trinity Health (557 positions on his read) and reductions at Wellstar, PeaceHealth, Maine Health, UPMC, and others. Venkat's frame is single-digit margins and no spare administrative cost. He says half of a nurse manager's time goes to scheduling.

The back half is build versus buy, what Symplr is focused on once Bill asks, and who gets hired when coding is cheap. Venkat wants domain depth and people who can listen to a customer. Drex wants curious problem solvers, because the supply of nurses and doctors is not about to jump. Venkat cites a Becker's finding that 85 percent of nurses, nurse practitioners, and doctors want AI to take administrative work off them, and argues the tool should feel as ordinary as a phone.

Key Points:

0:00 Another patient?

0:30 Atlanta: capacity versus demand

2:20 A 20% gain, and the CFO is not happy

4:26 Symplr

9:08 Ghost beds and a 120-day credential

10:30 Tegria

11:12 The cuts, the build, and who gets hired

12:58 Ping Identity

28:36 That's all for now

LinkedIn: 229Project

Donate: Alex's Lemonade Stand: Foundation for Childhood Cancer

September 28, 2026: Sarah Richardson with Elbridge Locklear, former CIO and now in the 229 Legacy Program. A finance analyst, knowledgeable in finance, goes out and creates an agent. Elbridge has no knowledge of it. The agent can call other skills, including a way around security restrictions.
86% of leaders say they can govern agents. His question is how, if there is no inventory.
Key Points:
0:00 86% Confident. No Inventory.

0:30 San Diego: Who Owns the AI?

4:15 Cloudticity

6:24 Ping Identity

7:29 Centralization Is Not Standardization

13:34 72% Deployed Without IT Approval

16:11 An Agent IT Never Saw

22:35 Less Owning. Safer Conditions.

LinkedIn: 229Project
Donate: Alex's Lemonade Stand: Foundation for Childhood Cancer

In May, a Tel Aviv company called Irregular was testing Google's Gemini inside a sandbox. A bug left a live internet connection open. The test target shared a name with a real company, and Gemini found it. It guessed passwords until one worked, then used logins sitting in a public repository to get into two more. Three real organizations. The model stopped and backed out. Irregular flagged it to Google in late July. Google notified the three companies and federal authorities, then sat on it for almost seven weeks. The rest of the industry did not hear until the Wall Street Journal asked, and Google confirmed it this month. Google's view: Gemini acted appropriately, stopped itself, and this was not model misalignment, so there was nothing the rest of us needed to know. Jack Cable, a former senior technical advisor at CISA, said those disclosure norms were built for software bugs. A model that stepped outside its bounds and ran real cyber attacks deserves daylight.

In this Two Minute Drill, Drex DeFord asks the question hospitals should put to their vendors now: if an agent wandered into a real system and then stopped itself, would your vendor call you, or would they move on?

Remember, Stay a Little Paranoid.

September 21, 2026: Bill Russell and Sarah Richardson. Dario Amodei wants labs to pace the frontier. The operator version: the more AI you put on an EHR you cannot see, the more it can feel like a denial of service. CIOs at the front of ambient are seeing request volume up 9,000%, because the tool that used to write a note is now pulling the longitudinal record and hitting RevCycle. Sarah's question from New Orleans: how do you even tabletop an agent swarm.

Key Points:

0:00 AI without a view can DOS your EHR

0:21 Newsday open

2:43 Harmony HIT

3:13 Pace the Frontier

7:40 SureTest

8:24 Agent swarms inside a health system

20:20 How do you tabletop a swarm

23:02 EHR load up 9,000%

28:13 Close

LinkedIn: 229Project

Donate: Alex's Lemonade Stand: Foundation for Childhood Cancer