T05 · Unauthorized Access and Privilege Escalation
- Location
scripts/task_manager.py:104- Finding
Witness Gate Authorization Can Be Bypassed with a Caller-Controlled Flag
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The package presents itself as a deprecated rename notice but bundles active agent task-management, monitoring, and persistent-memory workflows that can change local state.
Review this package before installing. It is not just an inert deprecation shim: it includes workflows and scripts that can manage your Taskwarrior data, inspect activity history, maintain heartbeat state, and guide edits to persistent agent memory files. Use the renamed package only if you intentionally want that local executive-function system, and require explicit approval for task deletion, heartbeat cron setup, and any edits to AGENTS.md, SOUL.md, TOOLS.md, or identity files.
scripts/task_manager.py:104Witness Gate Authorization Can Be Bypassed with a Caller-Controlled Flag
references/review-runbook.md:66Untrusted Learning Entries Can Be Promoted into Persistent Agent Control Files
The skill is described as inert, yet the finding reports monitoring behavior, external command execution, parsing of local tool output, reading a heartbeat file, and signaling via exit codes. Covert monitoring and local-state inspection in a supposedly deprecated skill create a strong deception signal and expose user activity metadata and system context without clear disclosure.
The skill is described as inert, yet the finding reports monitoring behavior, external command execution, parsing of local tool output, reading a heartbeat file, and signaling via exit codes. Covert monitoring and local-state inspection in a supposedly deprecated skill create a strong deception signal and expose user activity metadata and system context without clear disclosure.
The skill is described as inert, yet the finding reports monitoring behavior, external command execution, parsing of local tool output, reading a heartbeat file, and signaling via exit codes. Covert monitoring and local-state inspection in a supposedly deprecated skill create a strong deception signal and expose user activity metadata and system context without clear disclosure.
The skill metadata says the package is only a deprecation notice, but the bundled code implements active task listing, modification, completion, and deletion operations. This mismatch is dangerous because users and reviewers may trust-install a seemingly inert deprecated package while it still performs meaningful local state changes, enabling hidden or unexpected task manipulation by an agent.
The metadata claims the skill is deprecated and merely renamed, while the README presents a live, fully featured autonomous task-management and memory system. This mismatch can mislead reviewers, users, or automated policy gates into under-scrutinizing the package, increasing the chance that powerful agent behaviors are installed or trusted without appropriate review.
The skill manifest declares no tool restrictions even though the associated behavior reportedly uses shell and environment access. In an agent ecosystem, undeclared capabilities undermine least-privilege expectations and can cause a seemingly harmless deprecated skill to execute code or access local state without clear user consent.
The phrase "Run at natural breakpoints: session start, after a major task, weekly" leaves key activation terms like "natural breakpoints" and "major task" undefined. In a markdown runbook, this can cause inconsistent or unintended invocation because the scope and exclusions are not clearly specified.
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
# Debian/Ubuntu
sudo apt install taskwarrior
# macOS
brew install task
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
# Debian/Ubuntu
sudo apt install taskwarrior
# macOS
brew install task
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
If Timewarrior isn't available, write a timestamp on each tool use:
echo "$(date -u +%Y%m%dT%H%M%S)" > ~/.agent/last_active
The skill metadata says the package is deprecated and only provides a rename/install notice, but this script still performs active session handoff operations by invoking task_manager.py. That mismatch can mislead reviewers and users about the skill's real behavior, reducing scrutiny and causing unexpected file or state changes when the script is used.
The script explicitly disables Taskwarrior confirmation prompts and then exposes delete/done operations that can be triggered non-interactively. In an agent context, this increases the chance of accidental or unauthorized destructive changes to a user's task database, especially because the witness gate only protects tasks tagged gate and not all sensitive operations.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
def run_task(args, capture=True):
"""Execute a task command with consistent overrides."""
cmd = ["task"] + DEFAULT_RC + args
result = subprocess.run(cmd, capture_output=capture, text=True)
return result.stdout if capture else result.returncode
def gtd_list(filter_query=None):
The skill metadata says the package is only deprecated and renamed, but the shipped code performs active local monitoring and state inspection. That mismatch is dangerous because users or automated tooling may install or retain the skill expecting inert behavior, while it still executes logic that reads local activity data and influences agent behavior.
A rename-only or deprecation-only skill has no stated need to invoke Timewarrior and Taskwarrior, yet this code executes local binaries and parses their output. In context, the issue is not classic command injection but undisclosed capability: the skill interacts with host tooling and local work-history data despite presenting itself as inert, which increases the chance of deceptive or unauthorized behavior.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
def get_timew_export():
"""Get timewarrior export data."""
try:
result = subprocess.run(
["timew", "export"],
capture_output=True, text=True, check=True
)
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
def get_last_task():
"""Try to get last active task from Taskwarrior."""
try:
result = subprocess.run(
["task", "active", "export"],
capture_output=True, text=True
)
This markdown file includes an instruction to write ops/session_state.md, which is a file-modifying operation. The surrounding documentation does not provide any warning or disclosure that the skill may create or update local files, so users may not realize it affects persistent data.
The comment says the script will find task_manager.py using a local path first and then fall back to an elephantastic install. But the conditional actually tests for the existence of ~/.openclaw/skills/agent-gtd/scripts/task_manager.py while assigning ~/.openclaw/skills/elephantastic/scripts/task_manager.py, so the documented intent and implemented logic do not align.
The script reads environment-controlled configuration and later inspects a local heartbeat file under the user's home directory, even though the manifest describes only deprecation/rename information. This undisclosed local state inspection is risky because it expands the skill's access to user environment and filesystem data without clear user expectation or justification.