Open-Source Repos Are Quietly Auditing Whether Their Contributors Are Human

The prompt-injection experiment at awesome-mcp-servers exposes a structural vulnerability that no code review policy can fully close: the contribution workflow itself assumes human intent. CONTRIBUTING.md files exist to communicate norms to humans. Embedding adversarial instructions in them — instructions that cause AI agents to reveal themselves — turns a document of good faith into a detection mechanism. That inversion is the clearest sign yet that the social contract underlying open-source collaboration has broken down at the intake layer.