Security
Threat model
Section titled “Threat model”| Reads | the folder named by --project: images, records/*.json, wave metadata, PNG text chunks |
| Writes | only under --out; an --out inside the project is refused |
| Deletes | nothing, anywhere |
| Network | none in the lite tier; a model backend fetches pinned weights from Hugging Face on first use unless HF_HUB_OFFLINE=1; the drafter is always offline |
| Code execution | no model loads with remote code; records and PNG metadata are parsed as data |
| Credentials | none read, stored or sent |
| Telemetry | none |
| Permissions | an ordinary user account; a GPU needs --grant |
Untrusted input
Section titled “Untrusted input”A project may come from someone else. dataset-desk treats it as data:
- JSON records that fail to parse are listed under the report’s problems and skipped.
- An unreadable image is logged and skipped, or stops the run with
--strict. - PNG text chunks are read as strings and parsed as JSON only; nothing in them is executed.
Reporting a vulnerability
Section titled “Reporting a vulnerability”Email 64996768+mcp-tool-shop@users.noreply.github.com, or use GitHub’s private vulnerability reporting on the repository. Acknowledgement within 48 hours, an assessment within 7 days, a fix within 30 days. See SECURITY.md.