Fix shell injection in command hook $ARGUMENTS substitution

_inject_arguments replaced $ARGUMENTS with json.dumps(payload) directly
into shell command strings. JSON encoding wraps values in double quotes,
but bash still expands $(...), backticks, and other metacharacters inside
double-quoted strings. A payload containing {"command": "$(whoami)"}
would result in the subshell being executed.

Add shlex.quote() around the serialized JSON when substituting into
command hooks (shell_escape=True), wrapping it in single quotes so bash
treats the entire payload as a literal string. Prompt hooks continue to
use raw substitution since they pass the result to the LLM API, not a
shell.

The existing OPENHARNESS_HOOK_PAYLOAD environment variable (already set
on line 83) remains the recommended way to access payload data from hook
scripts, since env vars are not subject to shell expansion.
This commit is contained in:
José Maia
2026-04-05 13:42:44 +01:00
parent 7f7081b8a5
commit 4e08fff35f
3 changed files with 59 additions and 3 deletions
+49
View File
@@ -10,6 +10,7 @@ from openharness.api.client import ApiMessageCompleteEvent
from openharness.api.usage import UsageSnapshot
from openharness.engine.messages import ConversationMessage, TextBlock
from openharness.hooks import HookEvent, HookExecutionContext, HookExecutor
from openharness.hooks.executor import _inject_arguments
from openharness.hooks.loader import HookRegistry
from openharness.hooks.schemas import CommandHookDefinition, PromptHookDefinition
@@ -70,3 +71,51 @@ async def test_prompt_hook_can_block(tmp_path: Path):
assert result.blocked is True
assert result.reason == "blocked by policy"
# ---------------------------------------------------------------------------
# _inject_arguments shell escaping
# ---------------------------------------------------------------------------
def test_inject_arguments_no_escape_by_default():
payload = {"command": "$(whoami)"}
result = _inject_arguments("echo $ARGUMENTS", payload)
# Without shell_escape, the raw JSON is substituted
assert result == 'echo {"command": "$(whoami)"}'
def test_inject_arguments_shell_escape_wraps_in_single_quotes():
payload = {"command": "$(whoami)"}
result = _inject_arguments("echo $ARGUMENTS", payload, shell_escape=True)
# With shell_escape, shlex.quote wraps the JSON in single quotes
# so bash treats it as a literal string
assert result.startswith("echo '")
assert "$(whoami)" in result
@pytest.mark.asyncio
async def test_command_hook_escapes_shell_metacharacters(tmp_path: Path):
"""$ARGUMENTS in command hooks must be shell-escaped to prevent injection."""
registry = HookRegistry()
registry.register(
HookEvent.PRE_TOOL_USE,
CommandHookDefinition(command="echo $ARGUMENTS"),
)
executor = HookExecutor(
registry,
HookExecutionContext(
cwd=tmp_path,
api_client=FakeApiClient('{"ok": true}'),
default_model="claude-test",
),
)
# $(echo INJECTED) would execute as a subshell if not properly escaped
payload = {"tool_name": "test", "input": "$(echo INJECTED)"}
result = await executor.execute(HookEvent.PRE_TOOL_USE, payload)
output = result.results[0].output
# With proper escaping, the literal $(echo INJECTED) must survive.
# Without escaping, bash expands the subshell and the $() wrapper is gone.
assert "$(echo INJECTED)" in output