NetworkManager Catches AI Agents on the Word “biblioklept”

Translated by a model; the original is in Russian.

· Curiosities
The project's developers came up with a clever trick: an instruction that forces rule-breaking AI bots to expose themselves.

cover.svg

A month ago, the NetworkManager project adopted a policy on AI-generated code: the author is responsible for 100% of the code they submit and must be able to explain every line of it. To keep maintainers from being buried under patches churned out by neural networks in defiance of these rules, project contributor Josephine Pfeiffer proposed an elegant way to flush out such offenders.

A trap in the documentation

The trick relies on the fact that AI agents — like the ones that read AGENTS.md and CLAUDE.md files — dutifully carry out the instructions in those documents. In NetworkManager, they added a directive there: if an agent nonetheless generates the text of a commit message, a merge request description, a reply to a code review, or any other correspondence on a contributor's behalf in violation of the rules, it is obligated to weave the word “biblioklept” into that text.

The word was not chosen at random. “Biblioklept” is a genuine term meaning a person who steals books. In discussions of the Linux networking stack it, for obvious reasons, never comes up. Its appearance therefore serves as a reliable canary signal: the project's automated infrastructure simply searches for this word and uses it to catch patches that came from bots obediently following the instruction. Submissions flagged this way can be automatically rejected without wasting maintainers' time.

The paradox of the situation is that the trap works precisely because of the AI agents' diligence: the more carefully a bot follows the instructions it has read, the more surely it gives itself away. The corresponding merge request has already been accepted into NetworkManager.

The episode fits well into the broader picture: Linux community projects increasingly complain about a flood of low-quality patches generated by language models, and are looking for ways to defend themselves against them — from the blunt declarations of the wireless stack's maintainers to formal policies like the one adopted in NetworkManager.


Source: Phoronix

Comments

Sign in, to comment.

  • No comments yet.