If you found a security issue in faithful, please report it privately rather than filing a public issue.

• Email: mx@a9l.im
• GitHub: use private security advisories

Please include a description, steps to reproduce, and the version you are on. I'll respond in a few days and try to have a fix as soon as possible.

Only the latest release on PyPI gets security fixes. If you're on an older version, please upgrade.

A few things faithful does at startup touch your real environment.

• Reads config.toml from ~/.faithful/ (or $FAITHFUL_HOME). This file usually contains your Discord bot token and an LLM provider API key. Please keep it private.
• Connects to Discord with your bot token.
• Connects to your chosen LLM provider (OpenAI, Gemini, Anthropic, or any OpenAI-compatible host) with your API key.
• Reads and writes plain files in data/ for the example corpus. If [behavior] enable_memory is on, the memory tool can also create, edit, and delete plain files in data/memories/. Path traversal outside that directory is blocked.
• Makes outbound HTTP requests if [behavior] enable_web_search is on. For OpenAI, Gemini, and Anthropic this happens server-side on the provider's infrastructure. For OpenAI-compatible backends faithful itself queries DuckDuckGo and then fetches whatever URLs the model asks for.

If you don't want any of this, please leave enable_web_search and enable_memory set to false (the defaults).

• Rate limit incoming Discord messages.
• Restrict which URLs the model can fetch.
• Sandbox the LLM provider's responses.
• Vet who pings the bot.

All Discord slash commands (/upload, /add_message, /clear_messages, /memory, /status, /generate_test, and the rest) are restricted to the user IDs listed in [discord] admin_ids in your config.

Faithful sends your example messages, conversation history, and (if memory is on) the memory file contents to whichever LLM provider you point it at. Please only configure providers you are ok with sharing that with.