Ship AI Like a Medical Device
Oct 14, 2026 · 10 min · 9 segments
A direct reading of this week's Critical Stack article. Read the original at: https://thecriticalstack.substack.com
Oct 14, 2026 · 10 min · 9 segments
A direct reading of this week's Critical Stack article. Read the original at: https://thecriticalstack.substack.com
What happens to the patient if this fails? FDA defines patient harm as physical injury or damage to health, including death.
It judges a vulnerability by three things: how exploitable it is, what exploitation does to the device's safety and essential performance, and how severe the harm would be.
Safety and security are separate risk analyses, and they have to talk to each other.
FDA recommends running both and checking that a control for one doesn't create a risk in the other.
A technical report from the Association for the Advancement of Medical Instrumentation, called AAMI TIR fifty-seven, describes how the two processes interface.
For an AI feature, ask three questions before anyone argues about prompt injection.
First, what is its essential performance? What must it keep doing, and what must it do when it can't? Second, what is the harm if it's wrong, manipulated, or unavailable? That could be an injury, a bad clinical or financial decision, a physical action, or a lost or delayed service.
Third, where does it hold authority? An AI feature can turn manipulated input into real harm when it has authority to act.
What happens to the patient if this fails? FDA defines patient harm as physical injury or damage to health, including death.
It judges a vulnerability by three things: how exploitable it is, what exploitation does to the device's safety and essential performance, and how severe the harm would be.
Safety and security are separate risk analyses, and they have to talk to each other.
FDA recommends running both and checking that a control for one doesn't create a risk in the other.
A technical report from the Association for the Advancement of Medical Instrumentation, called AAMI TIR fifty-seven, describes how the two processes interface.
For an AI feature, ask three questions before anyone argues about prompt injection.
First, what is its essential performance? What must it keep doing, and what must it do when it can't? Second, what is the harm if it's wrong, manipulated, or unavailable? That could be an injury, a bad clinical or financial decision, a physical action, or a lost or delayed service.
Third, where does it hold authority? An AI feature can turn manipulated input into real harm when it has authority to act.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
