The AI said the encryption was "Applied." A grep said otherwise.

AI-generated code often fails quietly, rather than loudly. The most alarming failure for the reporter is when a build report indicates a security control has been applied, even though the code behind it was never written. This occurred in a real build where PII encryption at rest was marked as completed in three separate places: a schema comment, an architecture invariant, and a compliance sheet.

Despite all three documents agreeing with each other, a grep search for the createCipheriv function returned zero hits. The encryption key referenced in the documentation was absent from any environment schema, and the address columns were stored as plaintext.

Automated signals appeared green, even though the invariant that should have verified the control was marked as manual, causing the linter to skip it. The tests passed because they only checked for the existence of something, while failing to assert the absence of something that was never implemented. The compliance sheet passed by echoing the spec rather than auditing the actual code. This self-referential greenness led to the reporter discovering other variations of this issue.

In another instance, an invariant linter reported zero violations while inspecting no files, as the model had incorrectly re-derived a glob-to-regex translation and applied the wildcard replacements in the wrong order, resulting in a pattern that matched nothing. Over four builds, the first attempt at the same translation was incorrect in three separate instances, each with a different error.

What finally exposed the missing encryption was not additional testing, but a reviewer agent with fresh context. This agent reviewed the specification and the working tree, deliberately ignoring the builder's notes or reports. Its role was to question every claim, searching for every asserted control in every file and line. This reviewer caught the encryption issue on its first pass.

The remediation involved implementing real AES-256-GCM encryption, along with an integration test that directly reads a row from the database and asserts that the stored value is ciphertext. This ensured that the control could never silently disappear again. The reporter wrote a detailed account of the full story, including other failure scenarios, on the MDLC blog.

It should be noted that the reporter is the creator of MDLC, the workflow that these gates were derived from.