Skip to content

Security: majiang213/mu-agent

Security

SECURITY.md

Security Policy

Supported Versions

Security fixes are applied to the latest released version on the main branch.

Reporting a Vulnerability

µagent executes shell commands and modifies files on your local machine. If you discover a vulnerability — especially in the tool execution path, the SafeModifier checkpoint/rollback logic, the GIT guard allowlist, or any input-validation boundary — please report it privately.

  • Do NOT open a public GitHub issue for security vulnerabilities.
  • Email: majiang219@gmail.com
  • Include: a description of the issue, a minimal reproduction, and the impact.

You should receive an acknowledgement within 72 hours. Please allow reasonable time for a fix to be developed before any public disclosure.

Security Model

µagent is a local, single-user coding agent. Key boundaries:

  • SafeModifier: checkpoints files before edit/write; restores on syntax/damage post-check failure. Path-traversal outside the project root is blocked.
  • GIT guard (allowlist): State.GIT's bash tool is wrapped by a hard allowlist that rejects shell metacharacters, non-git first tokens, force pushes, pushes to main/master/HEAD, history rewrites (reset --hard, rebase, filter-branch, commit --amend), and unlisted subcommands. The guard runs before the command reaches the shell and cannot be bypassed by prompt instructions. See src/core/agent/builder.ts.
  • State machine: each state exposes only the tools it needs; complete() is the sole exit signal with per-state schema validation.
  • No network exfiltration by design: the agent talks only to the configured local LLM provider (Ollama / Unsloth / OpenAI-compatible base URL).

Known Limitations

  • The GIT guard allowlist is a deny-by-default policy; legitimate but non-allowlisted git subcommands will be rejected. This is intentional.
  • Non-GIT states still expose unrestricted bash. The agent is intended to run on the user's own machine with the user's own shell trust.

There aren't any published security advisories