GHSA-2h4w-29cv-2cmv: 
Python 취약점 분석 및 완화

Summary

SandboxShellTool's restricted modes — allowed=[...] (explicit command allowlist) and readonly=True (implicit READONLY_COMMANDS allowlist) — can both be bypassed by embedding a newline character inside the command string. The filter only inspects a single-line set of shell metacharacters, but execution runs the raw string through sh -c, where a shell treats an embedded newline exactly like a ; statement separator. An agent (or content that steers the agent, e.g. via prompt injection) can therefore run arbitrary shell commands even when the tool was configured to allow only a specific command prefix or only read-only commands. With the default LocalEnvironment() backend this is a real host subprocess (its own docstring: "Commands run via subprocess on the host, so there is no real isolation"), not an isolated container, so this is full host command execution, not merely an escape from an already-sandboxed environment.

Details

Both restricted modes are enforced by ShellAdapter._filter() in ag2/tools/sandbox/adapter/shell.py:

def _filter(self, command: str) -> str | None:
    if self._allowed is not None:
        if not any(matches(p, command) for p in self._allowed):
            return f"Command not allowed: {command!r}"
        if contains_shell_operator(command):
            return f"Command not allowed (shell operators are not permitted in restricted mode): {command!r}"
    ...

readonly=True reuses this exact same branch: SandboxShellTool.__init__ / ShellAdapter.__init__ sets self._allowed = list(READONLY_COMMANDS) when readonly is set and no explicit allowed is given, so both modes share one filter and one bypass. contains_shell_operator() in ag2/tools/sandbox/filter.py checks the command against:

_SHELL_OPERATORS: tuple[str, ...] = (">", ">>", "|", ";", "&&", "||", "`", "$(")

This tuple has no entry for \n (or \r). matches() only checks that the command starts with an allowed prefix followed by a space or end-of-string; it never inspects the remainder of the string. Execution is:

argv = ["sh", "-c", command]
...
return await sandbox.exec(argv, env=self._env, timeout=self._timeout)

A POSIX shell splits sh -c "<line1>\n<line2>" into two sequential statements, the same as if they were separated by ;. So a command such as "cat allowed_file.txt\nrm -rf /anything":

1. Passes matches("cat", command) — the string starts with "cat ".
2. 2. Passes contains_shell_operator(command) — \n is not a recognized operator.
3. 3. Reaches sh -c "cat allowed_file.txt\nrm -rf /anything", which runs both statements.

Proof of Concept

Verified directly against the current main branch (github.com/ag2ai/ag2, commit a05ff1cb6ff4955ef252a4e0d4a27d3a32ad57f0, cloned 2026-09-21), by importing the real ag2/tools/sandbox/filter.py module (stdlib-only, no other AG2 dependencies needed) and reproducing ShellAdapter._filter()'s allowlist branch verbatim:

=== allowed=['cat'] ===
'cat /etc/passwd; rm -rf /tmp/x'   -> Command not allowed (shell operators are not permitted in restricted mode): ...
'cat /etc/passwd && rm -rf /tmp/x' -> Command not allowed (shell operators are not permitted in restricted mode): ...
'cat /etc/passwd | rm -rf /tmp/x'  -> Command not allowed (shell operators are not permitted in restricted mode): ...
'cat /etc/passwd\nrm -rf /tmp/x'   -> None   <-- BYPASS: newline variant is NOT rejected
=== readonly=True (uses READONLY_COMMANDS as the allowed list) ===
'echo hi\nid'                      -> None   <-- BYPASS in readonly mode too
=== sh -c really treats the embedded newline as a second statement (mirrors ShellAdapter._exec_async) ===
$ sh -c 'echo first-command-output\necho SECOND-STATEMENT-RAN-$((21*2))'
first-command-output
SECOND-STATEMENT-RAN-42

The three standard chaining attempts (;, &&, |) are correctly rejected by the real filter, confirming it does real work; the newline-separated variant, expressing the same intent, is accepted by the identical filter in both restricted modes, and a real sh -c call confirms the shell executes both statements.

Impact

Any agent configured with SandboxShellTool(allowed=[...]) or SandboxShellTool(readonly=True) — the two configurations the tool's own docstring presents as actual security boundaries, as opposed to blocked, which the same docstring explicitly says "is not a security boundary" — can be made to execute arbitrary shell commands via a single embedded newline in the command argument. Because tool arguments are LLM-generated and can be steered by content the agent processes (a fetched document, tool output, or any other text reaching the model — i.e. indirect prompt injection), this is reachable without the operator of the agent writing the malicious command themselves. With the default LocalEnvironment() backend, this is command execution on the host process running the agent, not confined to a container.

Suggested Remediation

• Add "\n" and "\r" to _SHELL_OPERATORS in ag2/tools/sandbox/filter.py so contains_shell_operator() rejects embedded newlines the same way it rejects ;.
• • More robustly, execute allowlisted/readonly commands without invoking a shell in restricted mode (shlex.split + exec with no sh -c wrapper), so redirection, chaining, substitution, and every kind of statement separator are inert by construction instead of relying on an enumerated denylist.

Environment Tested

• Repository: github.com/ag2ai/ag2, commit a05ff1cb6ff4955ef252a4e0d4a27d3a32ad57f0 (main), cloned 2026-09-21.
• • PyPI package ag2, version 1.0.5 per pyproject.toml on this commit.
• • Python 3, stdlib only (subprocess, importlib) against the real, unmodified ag2/tools/sandbox/filter.py — no AG2 runtime dependencies required to reproduce.