Please report security issues privately via GitHub Private Vulnerability Reporting (Repo → Security → Report a vulnerability). Do not open a public issue for security-sensitive bugs.
A dedicated security email is intentionally not listed because none is operated for this project. GitHub Private Vulnerability Reporting is the channel; a maintainer must enable it in the repo's Security settings (see docs/maintainer-remote-setup.md).
Please include: affected commit/version, the command and config used, and a minimal reproduction. Acknowledgement is best-effort (this is a community project).
Only the latest release on main receives fixes. There are no LTS branches.
llama-serverhas no default strong authentication. The Quick Start binds to127.0.0.1(loopback). If you expose the server, put an authenticating reverse proxy in front and firewall it — never expose it to the public internet. An open server lets anyone run inference and exfiltrate model output.- Model output is untrusted. HunyuanOCR produces text/markdown from images; treat its output like any LLM output (may be wrong, may contain injected content from the source image). Do not execute or render it blindly.
- Self-hosted ROCm runner. The
rocm-runner-preflightworkflow runs on a self-hosted GPU runner and isworkflow_dispatchonly — it is never triggered by push or pull request, and must not be enabled on fork PRs. It checks out contributor code onto a machine holding weights/data; only a trusted maintainer dispatches it.
- Secrets, tokens, API keys, credentials.
- Model weights or quantized artifacts (
.gguf,.safetensors,.bin). - Private or copyrighted documents / datasets.
- Personally identifying information.
The .gitignore blocks common heavy-data extensions; if you accidentally commit
any of the above, notify a maintainer so it can be force-pushed off history.