Security fixes target the latest 2.1.x source and release.

Do not open a public issue containing tokens, credentials, session content, private paths, or exploit details. Contact the project owner at Discord: NN6 with a minimal description and arrange a private channel for sensitive evidence.

Include:

• affected version/commit;
• Windows version;
• runtime mode;
• reproduction steps with secrets removed;
• expected and actual security boundary.

• Local model and Kimi Web endpoints bind to loopback by default.
• Bearer tokens are random per launch and must not be logged/published.
• The desktop should stop only child processes it owns.
• AirLLM runtime loading must stay offline and target an existing local directory.
• Original Kimi needs normal internet access for login/search/fetch.
• Model outputs are untrusted text; local tool execution still requires the Kimi permission policy.

Do not expose the local server publicly, disable authentication, or run unknown model/code files with elevated privileges.

Before contributing, verify that the diff does not include:

• .env, tokens, API keys, cookies, OAuth files;
• data, isolation, Kimi config/credentials;
• session/wire JSONL;
• logs, screenshots, dumps;
• model weights, caches, runtime binaries.

Run:

git status --short
git grep -n -I -E 'ghp_|github_pat_|hf_[A-Za-z0-9]|sk-[A-Za-z0-9]'

This is a useful check, not a complete secret scanner.