Please report security problems privately, through GitHub: Report a vulnerability. Do not open a public issue.
Say what you found, how to reproduce it and which version you used. You will get an answer within a week. Once a fix is out, the advisory is published with credit to you, unless you prefer otherwise.
Only the latest release gets security fixes. The app updates itself, so staying on it is one click.
Botloft runs AI agents on your computer, with your account. Some things are by design and not vulnerabilities:
- Agents run as your user, without a sandbox. Keeping them apart is cooperative: an agent that may run any command (or one set to skip permission prompts) can read what your user can read, including other agents' folders and Botloft's own data. The app warns about this before you turn that on.
- The background service listens on
127.0.0.1only and asks every client for a token, which only your user can read. Other programs running as your user are trusted as much as you are.
A way around what an agent was allowed to do, a page or a message that gets an agent past a permission prompt, or a program reaching the service without the token is in scope and very welcome.
Antivirus detections of the releases are not security reports: see the code signing policy and open a normal issue.