Report vulnerabilities privately, never in a public issue:

• GitHub: open a private security advisory, or
• Email: security@runanywhere.ai.

Please include:

• what the issue is and what an attacker could do with it;
• the affected version or commit;
• steps to reproduce it, or a proof of concept;
• any suggested fix.

We acknowledge reports within 3 business days and keep you updated until the issue is fixed. Please give us a reasonable time to release a fix before you disclose it publicly.

Security fixes land on development and ship in the next release from main. Only the latest release gets them.

TrnRT is a model server with no authentication:

• The router listens on every interface by default (--host 0.0.0.0). Put it behind your own proxy or gateway, or pass --host 127.0.0.1.
• Under serve, engines listen on 127.0.0.1 only, so their /metrics endpoint is not exposed beyond the host. A standalone python -m trnrt engine binds --host (default 0.0.0.0).
• The tensor-parallel ranks talk over local TCP and shared memory. Keep ports 21000 and up closed to other hosts.
• The serving container runs with --privileged to reach the Neuron devices. Run it only on a host you control.
• Pin --revision when serving a Hugging Face repo id, so the weights and chat template cannot change under you.