No stable model has been promoted yet. Security fixes apply to the current main branch and the newest stable release once one exists.

Open a private GitHub security advisory at https://github.com/NT9712/tidelm/security/advisories/new. Do not include credentials, private data, or an exploitable model artifact in a public issue.

TideLM publishes model parameters only as SafeTensors. The loader sets trust_remote_code=False, pins the base revision, and requests SafeTensors. Never load untrusted pickle, .pt, or .bin files. Verify release assets with sha256sum -c CHECKSUMS.sha256 before loading them.

Generated code evaluation is opt-in. The harness rejects imports and dangerous AST nodes, applies process resource limits, and uses a timeout, but this is not a hardened sandbox. Run untrusted generations in a disposable container.

Secrets belong in environment variables or hosting-provider secret stores. Files matching .env*, .vercel/, model weights, and local blind-evaluation keys are ignored by Git.