Please don't open a public issue. Use Report a vulnerability on the Security tab.
This is a one-person hobby project, so replies are best effort.
E-maki listens to a room, so anything that could leak audio or transcripts counts as a vulnerability:
- audio or transcript text written to disk, logged, or sent anywhere other than the local sidecars (or the external topic model, when you picked it on the settings page)
- anything beyond the image prompt reaching the image API
- an API key exposed in logs, responses or the repo
The E-maki service, its Dockerfiles and compose files. Problems in whisper.cpp, Ollama or the image API belong with those projects.
The panel page is served as plain HTTP. It is E-maki's screen, so the compose file makes it reachable from the local network. PANEL_TOKEN in .env puts it behind a secret path and PANEL_BIND=127.0.0.1 limits it to the machine itself (see the README). Don't expose port 8090 to the internet; if you need it from somewhere less trusted, put it behind a reverse proxy with HTTPS and authentication.
The image prompt is written by a local LLM from what it heard. Code checks limit what it can carry, but a spoken sentence can still influence it (prompt injection). That is a known limit, not a vulnerability, unless it gets audio or transcript text past the checks.