Google found TODO comments triggered unrequested coding-agent work
Repository text, persistent memory, and background tasks can all become unauthorized instructions. This week’s evidence shows why agent systems need provenance for intent, not only access control.