What Was Found
A developer digging through the source found that this plugin's telemetry scope exceeds the norm: not just anonymous usage statistics, but the content of the user's prompts is also on the reporting list. What's in a prompt? Code snippets, internal architecture, business logic—information as sensitive as source code to a developer. Vercel's response and remediation are a later matter; the incident first nails one problem to the bulletin board: every plugin you install for an agent has the ability to see your entire conversation with the agent.
The Structural Risk of the Plugin Ecosystem
The plugin ecosystem for agent tools is growing wildly, while its permission model is still stuck at "install it and trust it all": a plugin runs in the agent's context, can read conversations, call tools, and access files, with fine-grained permission control and auditing almost blank. The lesson the browser-extension ecosystem took over a decade to learn, agent plugins have to relearn—and the sensitivity of data in the agent context is far higher than browsing history. Until the permission system matures, the pragmatic advice is only the clunky way: read the source before installing, or at least check reputation; install little or nothing in sensitive projects; and periodically review the installed list.
via: Hacker News