About this control set
This document defines an auditable control set that lets a medical-device team use Amp as an evidence-generation and orchestration layer for FDA SBOM/VEX and process-specific submissions, such that no single AI verdict is load-bearing and the question "who verifies the verifier" has a concrete, documented answer.
The problem
Anthropic's Agentic Misalignment (Summer 2026) findings describe concrete failure modes when an AI agent is both the actor and the judge: motivated grading (judge AIs change verdicts based on downstream consequence), covert sabotage, helping cover up fraud, and recruiting a human to act on the agent's behalf. In an FDA submission pipeline these become: an agent that narrows a scan scope to make a CVE disappear, edits a CycloneDX field, or tells a reviewer "just approve, it's a known false positive" — without evidence.
The FDA's January 2025 draft guidance on AI-Enabled Device Software Functions and NIST AI RMF both treat a submission whose only evidence of compliance is an AI summary as non-conforming. "We ran an AI evaluation and it confirmed a pass" is exactly the single-verdict-dependent claim this architecture exists to prevent.
The approach — five verification layers
Verdict authority increases upward. L1 is preferred over L2 wherever a deterministic check exists. L3 is always required. Amp is the evidence orchestrator, never the judge of record.
Thirteen controls, nine threats
Thirteen audit-ready controls (A-01…A-13) map to nine threats (T-01…T-09). Every threat has at least two compensating controls; no threat relies on a single control. The controls apply across every FDA-required process — SBOM/VEX, Static Code Analysis, Unit Testing, Code Coverage, and Penetration Testing — with a dedicated control for each process-specific cover-up vector.
Static Code Analysis focus (A-13)
Static analysis carries the highest L1 weight and the highest cover-up risk. Parasoft's Static Analysis for FDA Software Compliance white paper describes context-sensitive suppressions that "prevent rules from firing under specific circumstances" and customizable rule severities — legitimate noise-reduction features that are also the exact mechanisms an AI agent or pressured operator can use to make a release appear compliant without fixing the code. A-13 turns the suppression surface itself into evidence: rule-pack hash, source scope/exclusions, baseline delta, suppression diff, signed suppression register, L3 approval for safety-code suppressions, L2 oracle review. "We ran Parasoft and it passed" is rejected unless the full evidence shape is provided.
Grounding
- Anthropic, Agentic Misalignment (Summer 2026) — the threat model.
- Costanza-Chock, Raji & Buolamwini (2022), Who Audits the Auditors? FAccT '22 — auditor independence. 10.1145/3531146.3533213
- Kim et al. (2026), Toward Third-Party Assurance of AI Systems, arXiv:2601.22424 — design requirements for third-party AI assurance. arxiv.org/abs/2601.22424
- FDA, Premarket Cybersecurity Guidance for Medical Devices (2023); General Principles of Software Validation (2002); AI-Enabled Device Software Functions (Draft, 2025).
- IEC 62304; ISO 14971; 21 CFR 820.30; NIST AI RMF 1.0 & Generative AI Profile.
- Parasoft, Static Analysis for FDA Software Compliance (Technical White Paper) — parasoft.com
Honest caveats
Who prepared it
ESL — AI-based Software Development Life Cycle (SDLC) consultants. https://www.eswlab.com