Report vulnerabilities privately through GitHub Security Advisories, not public issues. Include the affected crate and version, steps to reproduce, and the impact. Expect an acknowledgement within 72 hours.
Only the latest release line (currently 1.0.x) receives fixes.
borg makes it easy to hand a model local resources. That's the point, and it's also the risk: treat everything a model sends to a tool as attacker-controlled (prompt injection).
In scope: a panic escaping the supervisor, cancellation that doesn't propagate, a budget that can be bypassed, a guarantee below that doesn't hold, or a release that ships a dependency with a known advisory.
TokioBackend runs ordinary subprocesses with your privileges: the child
can read and write every file you can and reach the network. To run
untrusted code, implement SandboxBackend over a real isolation layer:
- gVisor or Firecracker microVMs
- bubblewrap or nsjail
- hosted: E2B, Modal, Daytona, Cloudflare Sandbox
- agentjail (beta): lightweight Linux sandboxes for AI agents; check its own security notes before relying on it
What SubprocessTool and TokioBackend do guarantee:
- No inherited environment. The child gets only a passthrough list
(default
PATH,HOME,LANG,LC_ALL,TMPDIR,TERM, plus the Windows essentials). Your API keys don't reach it unless you pass them. - Model env is opt-in. The model's
envis ignored unless you allowlist keys, andPATH,LD_*,DYLD_*,BASH_ENV,NODE_OPTIONS,PYTHONPATH,GIT_SSH_COMMAND,GIT_CONFIG_*and similar are never accepted from it. - Exact allowlist.
SubprocessTool::builderrequires a program allowlist and a workspace root. Names resolve to absolute paths on your search path when the tool is built; anything else with a path separator is rejected. Allowing a shell or an interpreter still allows arbitrary code. - Workdir inside the root, after canonicalization. stdin is
/dev/null. - Process-group kill on Unix on timeout, cancel, drop, and when the direct child exits. On Windows only the direct child is killed.
- Limits: 30 s default timeout; output read in bounded chunks, 64 KiB per line and 1 MiB in total by default.
- Bad model input is a tool failure, never an error that aborts the run.
SubprocessTool::unrestricted drops the allowlist and workspace root; use
it only for trusted input. Most sandbox tests are Unix-only.
- An MCP config names an executable that runs as you. Never load configs you don't trust.
- Servers start from a cleared environment: a minimal passthrough
(
PATH,HOME,USER, locale and temp vars, plus the Windows essentials), names you add withpass_env, and the server's declaredenv(its${VAR}references resolve from your environment at load time).inherit_env(true)opts back into the full environment. - Tool names, descriptions, schemas and results come from the server and
reach the model verbatim, and
refresh()can change them (tool poisoning). Connect only trusted servers, allowlist tools, and gate side-effecting tools behind approval. - Each call has a timeout (default 60 s), honours cancellation, and has a result cap (default 1 MiB, applied after decoding).
Logs contain prompts, tool arguments and results. On Unix it creates
directories 0700 and files 0600 and refuses to write through a
symlink. It doesn't validate a directory that already exists, so point it
at a directory only you can write to.
The file tools in the docs are illustrative. Allowlist and normalize paths before writing anything to disk.