Security and privacy
What each user can see
Section titled “What each user can see”Reading records changes nothing, but it can reveal very sensitive information. So the DevKit decides what can be seen by how sensitive it is, and enforces this on the server, not in the AI assistant.
| Information | Who can see it |
|---|---|
| Practice scenarios, health-record checks | Anyone using the DevKit |
| Made-up records | Anyone, in the practice ABDM |
| Status of consents and fetches | Only the user who made them |
| Real record contents | Nobody through the DevKit, unless the operator declares a controlled setup for it |
| Aadhaar numbers and one-time passwords | Never shown to an AI assistant |
Safety rules
Section titled “Safety rules”- Your app decides who the user is, never the AI assistant.
- A person approves every request before it is sent, and the approval covers only that exact request.
- Record contents shown to an AI assistant are size-limited and marked as untrusted, so instructions hidden inside a record are not followed.
- The audit log never contains patient details or record contents.
- Passwords, keys and ABDM access tokens stay on the server.
Independent review
Section titled “Independent review”In October 2026 an independent review of the code found five problems, including two record deliveries arriving at once and records kept after consent was withdrawn. All five are fixed, each with an automated test that fails if the problem comes back.
Reporting a problem
Section titled “Reporting a problem”Please report security issues privately through the Security tab of the GitHub repository, not in a public issue.