Skip to content

Security: justinmclean/jumar

SECURITY.md

Security Policy

Reporting a vulnerability

Please report suspected vulnerabilities privately via GitHub Security Advisories — do not open a public issue. You should receive a response within a week.

Include the affected version or commit, a minimal reproduction, and any relevant logs or run journal excerpts.

Scope: what jumar promises (and what it does not)

Jumar dispatches agent processes to do real work on your machine, so its security posture is explicit and worth reading before you run it:

  • The enforced boundary is send, not fetch. Outbound-transmission vectors (mail, mailx, sendmail, ssmtp, msmtp, ssh, scp, sftp, rsync) are denied in every argv jumar dispatches, and git push / gh are hard-denied. Network fetch is allowed by default.
  • The Claude harness's shell is parsed, not pattern-matched. A fail-closed PreToolUse hook reads every Bash command the agent issues the way bash will and refuses git push and gh however they are spelled (control operators, quoting, substitutions, -c strings, heredocs fed to a shell, wrapper commands). A script file the agent writes and then runs is not read, and an interpreter can still open a socket.
  • MCP server inheritance is disabled for the Claude harness (--strict-mcp-config), so user-configured MCP servers cannot bypass the deny list.
  • The allow list is defence in depth, not a sandbox. With an interpreter allowed and the network reachable, the real control is running jumar inside a container/VM with restricted egress and no push credentials in the environment. Please run it that way.
  • Subtask environments are scrubbed (GITHUB_TOKEN, API keys, SSH agent socket removed), every subprocess is argv-only (shell=False), and verifier contexts are always fresh.

Reports that demonstrate a bypass of the send boundary, the argv deny list, the environment scrub, or the verifier-isolation guarantee are especially valuable.

Supported versions

Only the latest release receives fixes. Before the first release, fixes land on main.

There aren't any published security advisories