Act 33
The check that cannot fail
A vulnerability scanner and a test suite's own selftest were each built specifically to catch failure, and each had a blind spot exactly where it needed to see clearly. Both surfaced the same night.
That night asked the same question twice, at two different layers, before anyone noticed it was the same question.
A count that stood in for the truth
Section titled “A count that stood in for the truth”A vulnerability scanner had never had access to the container daemon on any host it ran against. Every scan of a floating tag was reading the registry, not the workload actually running: a plausible count standing in for the real one, and nothing in the result said which was which. One host got a read-only proxy set in front of the socket instead.
A second scanner, and what it caught
Section titled “A second scanner, and what it caught”A second scanner went in alongside the first after the first turned out to have its own version-specific blind spot: clean on one base image version, while the second scanner, run against the identical image, came back with eight real criticals. After that, each finding’s report started carrying whether it was actually fixable, not just how severe it was, because severity alone had been training people to tune the whole channel out. Before anything merged, one more catch turned up: the second scanner had no cache volume of its own, so it was silently pulling down its entire vulnerability database again for every image it scanned. Repeated dozens of times over, that 80-second cost sat inside a test run that took 30 minutes, long enough that it couldn’t be missed.
Then the same pattern turned up one layer above, each catch leading to the next. Both catches would turn out to rest on one point by the night’s end: a check that cannot fail by its own construction is indistinguishable from a check that has simply never been asked the question able to break it, and nothing in the check itself says which one it is. A check can’t fail the way the bug does carries where that point goes.
An exit code standing in for three different answers
Section titled “An exit code standing in for three different answers”A probe suite existed for exactly one reason: so that getting nothing back could never read as a passing check. It still had a blind spot, and the blind spot was in its own selftest. One step, the one that lists files over a remote connection, called on privilege escalation without checking whether it would be granted, assuming it always would be. On the host where that escalation is actually refused, three different situations, a file that was missing, one that was denied, and one that was perfectly readable, all failed in exactly the same way, and the suite counted itself passed on nothing more than a matching exit code of 2.
The mirror mistake
Section titled “The mirror mistake”That was fixed. What followed right behind it was the same mistake in reverse, caught in the other work session that night: the cases testing for a missing or denied file had been written as hard assertions instead of skips, and on the two hosts where that kind of passwordless escalation is genuinely granted, they were failing the suite outright, falsely. That mirrored the rule stated the other way around, that “skipped is not passed.” That same mirror seen from the other side, “unreachable is not failed either,” was why those cases were turned into skips instead.
Cannot connect to the Docker daemon