A guided walk from the AI vulnerability explosion to the AI response: premarket testing, threat-mitigation V&V, exploitability verification, and the evidence FDA reviewers actually want. Step through it below, or open it full screen.
Enter your details and the full explainer opens right here, no download required.
The explainer is a guided walk from the AI vulnerability explosion to the response: premarket testing under the Secure Product Development Framework, threat mitigation V&V, exploitability verification, and the evidence FDA reviewers want. Access is free. Enter your details and it opens in place, with a full screen view if you prefer.
FDA premarket cybersecurity guidance expects testing to live inside the Secure Product Development Framework, which means periodic testing through development rather than one engagement before submission. Where a report has no open findings, it still has to prove you looked, with an enumerated list of every test case launched against the product scope. A clean report with no testing history reads as untested.
Traditional V&V asks whether a requirement is present. A login screen exists, so the authentication requirement passes and nothing was attacked. Threat mitigation testing attacks the implementation instead: credential stuffing, session fixation, lockout bypass and token replay against that same login. FDA asks for proof the mitigation works, not proof the requirement was written down.
Threat model drift is what happens when cybersecurity testing finds vulnerabilities in places the threat model says nothing about. Those two artifacts are supposed to move together, and a reviewer can see when they do not. The fix is resolution: expand the model to process level detail so the model and the testing stay in sync.
An untriaged finding keeps the default rating it arrived with, and against cybersecurity patient safety criteria that often screens as uncontrolled risk. FDA can ask for the last year of vulnerability data and triage on specific products and releases. That the true severity was low does not repair a record showing it was missed.