TSFM-Lab is a local tool: it runs on your own machine, listens on 127.0.0.1 by default and sends nothing anywhere except the downloads you ask for (model weights and public datasets from Hugging Face, pinned to a commit and checked against a SHA-256). The page Security model says what it protects against and, as plainly, what it does not (for example, there is no login when you use --tailscale).

Please do not open a public issue for a security problem. Use GitHub's private reporting instead: the repository's Security tab, then Report a vulnerability. Say what you found, how to reproduce it and which version (git rev-parse --short HEAD) you ran. You will get an answer as soon as the maintainer reads it; this is a one-person project, so there is no response-time guarantee.

• Anything that lets a web page, another local user or a file the app reads run code, read files outside the project or reach the server from another host (the Host/Origin checks, the upload and download paths, the run and dataset identifiers taken from URLs).
• A weight or dataset file that gets used without being checked (the safetensors-only rule, the pinned commits and checksums).

GitHub's Dependabot alerts are on, and the pinned versions in uv.lock are also checked against the OSV database with pip-audit (last on 30 October 2026, for the CPU, CUDA and macOS configurations): no known vulnerability is open. On that date the lock file moved every platform to PyTorch 2.13.0 and setuptools 84.2.1, which closed GHSA-rrmf-rvhw-rf47 (torch, torch.jit.script memory corruption, fixed in 2.13.0) and GHSA-xmj3-9gf7-h52w, GHSA-h35f-9h28-mq5c and PYSEC-2026-3447 (setuptools, fixed in 83 and 84.1). The app used none of the affected parts. pip-audit cannot look up the local builds torch 2.13.0+cpu and +cu130 (they are not on PyPI); it checks the same release on the macOS configuration and Dependabot reports all alerts as fixed.

The OSV audit is not run automatically: repeat it after updating the lock file (uv export --no-hashes -o requirements.txt, then uvx pip-audit -r requirements.txt --no-deps).

Out of scope: vulnerabilities in the libraries the app depends on (report them upstream; uv.lock pins the tested versions), and the models' own behaviour (what they forecast, their licences, their training data).