@hulumi/drift: Drift classifier fails open on adapter errors and over-promotes Mixed verdicts

Affected: @hulumi/drift < 1.4.0 — Fixed in: 1.4.0 — Severity: Medium — CWE-755 (Improper Handling of Exceptional Conditions)

Summary

@hulumi/drift runs four adapters that each ask a different question about whether a resource has drifted (Pulumi-state diff, provider-version change, CloudTrail event, etc.). A classifier combines the adapters' answers into a verdict like None / none, ConsoleBreakGlass / high, or Mixed / high, and caches the verdict for 6 hours by default.

Two related bugs from one root cause — the classifier only read each adapter's detected: true/false field and ignored whether the adapter itself succeeded:

1. Cached "all clear" on adapter failure. When an adapter failed (e.g. transient network error from the Automation API), the classifier read detected: false, concluded "no drift", and cached the verdict as None / none for 6 hours. A single transient failure could mask real console-break-glass mutations for the rest of the window.
2. Mixed verdicts without real evidence. The Mixed / high and ConsoleBreakGlass / high verdicts (incident severity) could fire on the "the CloudTrail probe round-tripped successfully" signal rather than actual evidence that anything had been changed via the console. Normal provider-API churn could end up falsely escalated to incident severity.

Impact

Consumers running drift detection in CI / cron could see transient adapter failures silently cached as "all clear" — masking real attacks for up to six hours — or see ordinary provider-version churn falsely promoted to incident severity. Either way, the verdict source was unreliable for downstream incident workflows that gate on it.

Patches

Upgrade to @hulumi/[email protected]. Classifier-only fix (the TLA+-verified 6-row verdict matrix is byte-identical):

• adapter failures now fail closed to Unknown / low, and degraded verdicts are not written to the cache;
• the Mixed / ConsoleBreakGlass promotion now requires real CloudTrail event evidence rather than probe liveness.

Workarounds

Setting options.minConfidence: "medium" on the classifier call prevents the degraded None / none from being cached (it doesn't meet the threshold), partially mitigating case (1). No workaround for case (2).

Resources

• PR #178 (Cluster D); regression tests in packages/drift/tests/classifier-fail-closed.test.ts.