← Insights

When a security control can hurt the patient

Most security controls fail closed by default. Deny, lock, disconnect.

In a medical device, that default can be the hazard.

Three terms get mixed up here:

  • Fail-secure: when something breaks, deny access.
  • Fail-safe: when something breaks, stop and enter a safe state.
  • Fail-operational: when something breaks, keep delivering therapy safely, possibly at reduced performance.

They don't always agree. A control that fails secure can force a stop when the right answer is to keep going. Common places it shows up:

  • A session timeout that locks the controls mid-procedure.
  • An expired certificate that cuts a connection a clinical function depends on.
  • A forced update that restarts the unit while it is in use.
  • An account lockout when someone needs access urgently.

How I approach it:

  1. Decide per function, not per product. List what the device must keep doing with no network, no cloud and no identity service, what it can keep doing in a degraded mode, and what it is allowed to refuse or stop.
  2. Keep therapy-critical function independent of remote security services. Authorize locally against cached policy where appropriate, and degrade remote features first.
  3. Record the decision in both risk analyses. The security risk file and the ISO 14971 risk file, cross-referenced. A control that introduces a hazard is a risk in its own right, with its own mitigation.
  4. Test the failure paths. Expired certificate, unreachable identity provider, full disk during an update. Keep the evidence that it fails the way you decided.

There is no universal answer. Some functions should stop, some should keep running degraded, and remote access to a device that delivers therapy should fail closed. The point is to make the call at design time, from requirements, and write it down, rather than inherit whatever the framework defaults to.

The test: pick one security control. What does the patient experience when it fails?

If nobody can answer, the control has not been fully designed.

Seen this differently?

Questions, corrections and counterexamples are welcome.

Related

Keep reading