Tacitus reads mail and calendars and drafts in the user's name. Its design limits what the model can do without the user's click; docs/SECURITY.md has the threat model, the five layers and what a prompt injection cannot do.
Report privately through a security advisory. Never open a public issue for a vulnerability. Expect a first reply within seven days; the project has one maintainer.
A report helps most when it says which layer it crosses: a way to send
mail, create or change an event, or delete, without the user's click; a
way for the ea-bot user to read the owner's tokens or private
directory; a way for mail content to reach the model when the mail gate
keeps it; or a way for a plugin to act beyond the capabilities it
declared.
The main branch and the latest tagged release. Older tags get no fixes.
The model writing a poor draft, which the user reviews before it is sent; a reporter's own machine, chat server or tokens being exposed by their configuration; and Claude Code's own behaviour, which belongs to Anthropic's program.