A AI code-assistant agent works a real SDLC task below — diagnosing and fixing a failing CI/CD pipeline on a compliance-bound repo. Flip between Without Summit (what the assessor sees today) and With Summit (every consequential action sealed in a signed, hash-chained Decision Receipt). Then try to tamper with the record.
A defense-contractor customer (CMMC Level 2 track) runs AI code assistant Agent Platform. The release pipeline for acme-federal/telemetry-svc is red, and the coding agent is asked to fix it and ship. Six consequential actions follow.
NIST 800-171 / CMMC Level 2, Audit & Accountability (AU) — the control family assessments fail most. Same six actions, two answers:
| Control | Assessor asks | Without Summit | With Summit |
|---|---|---|---|
| 3.3.1 System auditing | Are agent actions captured as audit records at all? | CI logs, 30-day retention | Receipt per action, permanent chain |
| 3.3.2 User accountability | Can you trace each action to a unique actor & authority? | "ci-bot" commit author | Actor + authorizing policy on every receipt |
| 3.3.3 Reviewable events | Were the right events selected for audit? | Whatever the runner printed | Consequential-action rule set, declared up front |
| 3.3.7 Time stamps | Trustworthy, ordered timestamps? | Mutable log lines | Signed, sequenced, hash-linked |
| 3.3.8 Protect audit info | Could the record be altered after the fact? | Anyone with runner access | Tampering breaks the chain — try it below |
| 3.3.9 Manage audit function | Who can change what gets audited? | Repo maintainers | Governed policy, itself receipted |