Answer Key — Tool Audit
Admin / self-check reference
The point is not matching this table row-for-row. The point is that your own table uses the same seven columns, forces the same questions, and drives a clear Verdict for every tool you run.
Worked reference
| Tool | Reads | Writes | Net | Exec | Telemetry | Verdict |
|---|---|---|---|---|---|---|
| Cursor (IDE) | Whole repo buffers | Files in open workspace | AI provider APIs | Yes via agent mode | On by default | Restrict (disable agent mode for sensitive repos, disable telemetry) |
| Claude Code CLI | cwd and below | cwd | Anthropic API | Yes (shell via Bash tool) | Minimal | Keep with hooks and permission modes |
| GitHub Copilot | Whole open file + nearby | No | Microsoft / OpenAI | No | On | Restrict (disable on private repos, disable telemetry) |
| Local MCP filesystem server | Mounted paths | Same mounted paths | None | No | None | Keep, scope the mount narrowly |
| Grammarly browser extension | Every textarea on every site | Rewrites text in place | Grammarly cloud | No | On | Remove from work profile |
| Ollama (local model runner) | Prompt input only | Model cache | Ollama registry for pulls + update checks | No | Auto-update ping | Keep, disable auto-update, pin versions |
Note: Your own audit will differ. The point is having the table and reasoning through each column.
What to look for in your own table
- No blank cells. If a cell is blank, the tool has not actually been audited — that is a research task, not a completed row.
- Verdicts that match the row. A tool that reads your whole repo, writes files, hits a remote API, runs shell commands, and has telemetry on by default cannot honestly end up as "Keep" with no further action. Either the Verdict is wrong or the mitigations belong in the cell.
- All five categories covered. If your table has no browser extensions in it, double-check: it is the most-forgotten category and almost nobody actually has zero AI browser extensions installed.
- Update channel noted somewhere. If the tool auto-updates and you have not pinned it, your Verdict should reflect the supply-chain exposure, not just today's behavior.