
The healthcare software failures that cost the most aren’t caused by missing policies. They come from design choices made in year one: how one customer’s data is kept apart from another’s, what the system records about who touched what, where patient data lives, and how much each credential can reach. A policy can be rewritten in a week. Those four decisions take a rebuild to change. So treat the auditor as a user of the system, with requirements you design for from the first sprint.
What the enforcement record says
From January to June 2013, an employee at Montefiore Medical Center in New York accessed the records of 12,517 patients and sold their information to an identity theft ring. The hospital found out in May 2015, when the New York Police Department told it (BankInfoSecurity). The resolution agreement with the HHS Office for Civil Rights (OCR) cites three potential violations: no accurate risk analysis, no procedures to regularly review system activity, and no mechanisms to record and examine activity in all systems holding patient data. Montefiore paid $4.75 million. In OCR’s words, the hospital was unable to prevent the attack “or even detect the attack had happened until years later.”
It isn’t an outlier. At Northeast Radiology, unauthorized individuals accessed an imaging server from April 2019 to January 2020, exposing 298,532 patients. Its corrective plan includes a written process to review audit logs and access reports (HHS). In March 2026, OCR settled with MMG Fusion, a software company whose 2020 breach affected about 15 million people. OCR opened that investigation in 2023, after a complaint and patient data posted on the dark web (HHS).
The corrective plans are mostly engineering. OCR’s standard advice opens with “Identify where ePHI is located in the organization, including how ePHI enters, flows through, and leaves.” That’s an architecture question, not a paperwork one.
The rule already describes a system
The HIPAA technical safeguards in 45 CFR 164.312 read like a spec. Allow access only to “persons or software programs” that have been granted it. Give every user a unique identity. Implement “mechanisms that record and examine activity” in every system that holds patient data. “Software programs” reads differently now that agents call tools.
The proposed update to the Security Rule goes further: a technology asset inventory and a network map of how patient data moves, refreshed at least every 12 months, plus network segmentation and multi-factor authentication. It’s still a proposal, and the federal regulatory agenda lists final action for July 2027. That map is only cheap if your architecture makes it obvious.
Four decisions you make in year one
How tenants are separated. Take a hypothetical clinic-scheduling platform that serves a few dozen clinics from one database, with a clinic ID column that application code is supposed to filter on. One new reporting endpoint forgets the filter, and one clinic sees another clinic’s patients. A policy can’t prevent that. Separation enforced in the data layer, through row-level security or separate schemas and keys, can. Retrofitting it later means migrating every table.
What gets logged. Most systems log errors. Auditors ask about access. For every read of a patient record, record the real person, any software acting for them, which record, which fields, and when, in an append-only store. Don’t let the log become a second copy of the chart. The test is the auditor’s question one security guide uses: “show me all access to Patient X’s record between March 10 and March 17” (Aptible). If that takes a week, you’re running Montefiore’s setup. As I’ve written before, access control without an audit trail is half a solution.
Where data lives. Patient data leaks into error trackers, analytics tools, support inboxes, backups and now model providers. Each must be in the risk analysis, and is often a vendor needing a business associate agreement. I’ve argued that your data flow diagram should exist before your first prompt template does. Drawing it early doesn’t slow shipping. It moves the pain to when it’s cheap to fix.
How access is scoped and revoked. That means one identity per person and per service, no shared admin accounts, short-lived credentials limited to what the task needs, and revocation as a single, tested step. The proposed rule would require notifying certain regulated entities within 24 hours when a workforce member’s access changes or ends. The platforms that win will make auth, scopes, and revocation boring.
Agents make this urgent
Agents multiply these surfaces. AWS’s reference architecture for HIPAA agents describes patient data entering through the prompt, flowing through tool calls, sitting in agent memory and trailing through logs. In its prior-authorization example, an eligibility tool returns the member’s full record, Social Security number included, when the task needed coverage status (AWS). Its point: audit logs capture database queries, “not the cascade of model inputs and outputs.”
Treat agents like untrusted operators: narrow scopes, explicit approval for irreversible actions, and logs you can audit. Every tool call should carry the identity of the human it acts for, and tools should return only the fields the task needs. Telling the model to be careful while the tool returns the whole chart controls nothing.
Write the auditor’s user stories
The cheapest compliance work there is: a few backlog items for a user most teams forget.
- As an auditor, I can list every person and program that read a given patient’s record in a date range, in minutes.
- As an auditor, I can see every system that holds patient data from a map generated from infrastructure, not drawn by hand.
- As an auditor, I can confirm a departed employee’s access was revoked everywhere, with a timestamp.
- As an auditor, I can show that one customer’s query cannot return another customer’s records.
Build them in the same sprints as the clinician’s stories. Policies describe what you intend. Architecture decides what you can prove.
About Ganesh Kompella
Ganesh Kompella is the founder of Kompella Technologies, a fractional CTO and CPO firm working with healthcare, fintech, and SaaS startups from pre-seed through Series B. 15+ years and 75+ products shipped, $140M+ ARR built, one IPO guided.
Sources
1. Marianne Kolbasuk McGee. Medical Center Fined $4.75M in Insider ID Theft Incident. BankInfoSecurity, 6 February 2024.
2. HHS Office for Civil Rights. Resolution Agreement and Corrective Action Plan with Montefiore Medical Center. HHS.gov, November 2023.
3. HHS Office for Civil Rights. HHS Office for Civil Rights Settles HIPAA Security Rule Investigation with Northeast Radiology. HHS.gov, 10 April 2025.
4. HHS Office for Civil Rights. HHS’ Office for Civil Rights Settles HIPAA Investigation of MMG Fusion, LLC Breach Affecting 15 Million Individuals. HHS.gov, 5 March 2026.
5. 45 CFR § 164.312. Technical safeguards. Legal Information Institute, Cornell Law School.
6. HHS Office for Civil Rights. HIPAA Security Rule Notice of Proposed Rulemaking: Fact Sheet. HHS.gov, 27 December 2024.
7. Office of Information and Regulatory Affairs. RIN 0945-AA22, HIPAA Security Rule to Strengthen the Cybersecurity of Electronic Protected Health Information. Unified Agenda, 2026.
8. Mat Steinlin. Agentic AI in healthcare: how to secure LLMs with tool access. Aptible, April 2026.
9. Ganesh Kompella. “Access control without an audit trail is half a solution”. X, 15 September 2026.
10. Ganesh Kompella. On privacy review for AI agents and data flow diagrams. X, 29 September 2026.
11. Ganesh Kompella. “Make auth, scopes, and revocation boring”. X, 1 October 2026.
12. Ray Chang and Vignesh Srinivasan. Architecting HIPAA-compliant AI agents to safeguard health data with AWS. AWS Public Sector Blog, 14 August 2026.
13. Ganesh Kompella. “Treat agents like untrusted operators”. X, 25 September 2026.

