Email info@neatlabs.ai with "DOCKETLAB security" in the subject. Please don't open a public issue.

Include what you can: the version, the affected component, and a way to reproduce it. We'll acknowledge receipt and keep you updated on the fix.

DOCKETLAB runs locally and holds two things worth protecting: your API keys and whatever corpus you've pulled.

• Keys are stored in settings.json under your data directory, outside the source tree, in plaintext. They are never logged, and the browser only ever receives the last four characters. Anyone with read access to your user profile can read them — treat that file the way you'd treat any credential on disk.
• Nothing is transmitted anywhere except to the APIs you configure: regulations.gov, the Federal Register, and Anthropic. There is no telemetry and no server component.
• The console binds to 127.0.0.1 only. Do not expose it to a network — it has no authentication because it is not designed to have any remote users.

Comment text is arbitrary public input from anyone who filed on a docket, and it flows into a language model prompt. Treat model output as untrusted: the extraction stage validates provisions against a strict format and discards anything unrecognized, JSON is parsed defensively, and a malformed response skips a unit rather than failing a run. Attachment downloads are streamed with a size ceiling so a single oversized file can't stall a pull.

If you find a way to make crafted comment text change tool behaviour beyond its own analysis row, that is a vulnerability and we want to hear about it.