Please report security issues privately, through GitHub's private vulnerability reporting: open the repository's Security tab and choose Report a vulnerability. Do not open a public issue, pull request or discussion for a security problem.

Include what you found, how to reproduce it, the inferencekit version (inferencekit --version) and the impact you expect. We aim to acknowledge reports within 5 working days and to agree a disclosure date with you once a fix is ready.

Only the latest released version receives security fixes.

Performance work produces files that can leak secrets. Treat them as sensitive.

• Nsight Systems (.nsys-rep, and the .sqlite exported from it) and Nsight Compute (.ncu-rep) files can embed the profiled process's environment variables, command lines, host names, user names and file paths. An environment variable can hold an API key or an access token (for example a Hugging Face token).
• Engine logs (vLLM, SGLang, llama.cpp, TensorRT-LLM and custom engines) can print their configuration, environment, command line, request contents and paths.
• Evidence records under runs/ store command lines and the resolved engine configuration.

What inferencekit does about it:

• inferencekit nsys capture and inferencekit ncu capture run the profiled program under env -i, with an allowlist of environment variables (PATH, HOME, CUDA_VISIBLE_DEVICES, plus the names you pass with --keep-env). They refuse to pass names that look like secrets (names containing TOKEN, SECRET, PASSW, KEY, CREDENTIAL, AUTH, COOKIE and similar), and record only the names of the variables passed, never their values.
• runs/, *.nsys-rep, *.ncu-rep and *.sqlite are listed in .gitignore.

What you should do:

• Profile only through inferencekit nsys capture / inferencekit ncu capture, not by calling nsys or ncu directly with your full environment.
• Never publish raw profiles or engine logs without scanning them first for tokens, keys, host names, IP addresses, user names and private paths. Even a capture made under env -i still contains command lines, host names and paths.
• Share derived summaries (inferencekit nsys analyze, inferencekit ncu summarize, evidence records) instead of raw profile files where you can, and review those too before publishing.
• Never pass a secret on a command line; read it from an environment variable by name.