Skip to content

Security: Barman-Zarei/Brok

Security

docs/security.md

Security model

Tool Risk Confirmation
read/list/search, git status/diff/log/branch, inspect LOW none
write/create/move, git checkout/stash, run_tests/linter/formatter, fetch_url MEDIUM configurable
delete_file, git_commit, github_create_* HIGH required
git_push, sudo/system/remote-changing commands CRITICAL always (no setting can skip)
dangerous commands (rm -rf /, mkfs, dd to disk, curl|sh, credential files, fork bomb…) BLOCKED refused before any prompt
  • Workspace jail: paths are resolved with realpath; escaping the project (.., absolute, symlink) is refused.
  • Secrets files (.env, *.pem, *.key, ssh keys) cannot be read or written by the agent; git_commit never stages them.
  • Commands run without a shell (argv), with API-key/token env vars scrubbed, with a timeout and output cap.
  • Git refs are validated (no option injection). fetch_url rejects non-http(s)/private/loopback addresses (SSRF).
  • Keys come from env vars or the OS keyring only; never from config, logs or memory (redact() masks them; MemoryStore refuses text that looks like a secret).
  • Plugins cannot lower risk: their tools are floored at MEDIUM/configurable.
  • The .git directory is off-limits to file tools (a planted core.fsmonitor or hook would run code on the next git status), and git itself runs with core.fsmonitor=false; delete/move act on a symlink itself, never its target.
  • fetch_url also re-checks every redirect and the address it actually connected to (SSRF / DNS rebinding); the GitHub and character-shop clients only send a token to their own origin and never re-send it on redirects.
  • A running command is killed together with everything it spawned when you press Stop or the timeout hits.
  • Config files (config.ini, brok.json, memory) are written atomically and are private from the first byte.
  • health findings are heuristic, not a security audit.

Limits you should know: command classification is pattern-based (it can be fooled by obfuscation, which is why anything unlisted asks first); there is no OS-level sandbox/container yet.

There aren't any published security advisories