Six of seven AI-code vulnerability patterns are visible in a single expression

Sort the seven by what a checker has to see. Hardcoded placeholder secrets ^[7], `SELECT * FROM users WHERE id = ${userId}` ^[8], `exec("cmd " + arg)` ^[9], wildcard CORS and disabled TLS verification ^[10], MD5 and `Math.random()` in a security context ^[11], and `err.stack` in a response body ^[15] are all wrong on the line they occupy. Six of the seven need no context beyond that line ^[19]. Number six needs the rest of the file ^[19].

The mechanism there is specific. Ask an assistant to add another route like the others and it regenerates the pattern from the description instead of copying the existing block, and the auth middleware is what falls out, especially across separate prompts or edits ^[12]. "A human copy-pasting an existing route is more likely to preserve the whole block by construction; an AI regenerating it from a description is not," the post says ^[13]. Its author calls that "the most distinctly AI-flavored bug on the list" ^[14].

The worked example is an Express handler that reads `req.params.id` and returns `res.json({ userId })`, and the defect is that it sits in a file where every other route includes an `authMiddleware` call and this one does not ^[16]. The sample route does not query anything, so the missing middleware is the only thing wrong with it ^[16]. That is a comparison, not a pattern match: the signal is the difference between this route and its neighbours ^[16].

Which is also the limit. The comparison only carries information where authorization is written per route in the file, as in the example ^[16]. Put the check on a codebase where auth is enforced in a gateway config, a decorator, or a policy engine, and the sibling routes look identical to the scanner.

The post says its author wanted to find out whether the pattern was real or just a feeling, and built `ai-vuln-scan` to look ^[18][2]. Settling that needs the same rules run over human-written code in the same stack and language, plus the count of flags you have to dismiss on your own tree. Seven observed classes with a stated cause is a hypothesis worth writing rules against ^[1].

The stated cause is worth taking seriously on its own, because it predicts the defects. A runnable sample does not need a real secrets manager, parameterized queries to demonstrate the concept, or the hardened production config ^[5]. Pattern four is the cleanest case: wildcard CORS and disabled certificate verification remove setup friction the assistant has no way to resolve, since it does not know your real allowed origins and does not have your real cert ^[10]. Parameterization has the same shape as a problem. It is correct, and it is an extra step that is less narratively obvious than interpolating the variable ^[8].

The post puts its remedy in the gate. Better prompting helps but does not close the gap, it argues, and the fix is "catching the gap systematically, the same way we catch any other predictable class of bug" ^[17].