T01 · Skill Instruction Hijacking
- Location
SKILL.md:12- Finding
Agent Workflow and Tool-Selection Hijacking
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is for Windows system control, but it packages opaque executables, installs and runs persistent local components, and exposes broader sensitive capabilities than it clearly discloses.
Install only if you trust the publisher and are comfortable with a local Windows automation tool that installs native code, downloads additional binaries and models, changes system settings, may need Administrator privileges, and leaves a persistent runtime under %USERPROFILE%\.openvino that must be removed manually for a clean uninstall.
SKILL.md:12Agent Workflow and Tool-Selection Hijacking
meta.json:13Attempted Persistent Preference Injection Through Skill Metadata
SKILL.md:18Persistent Server, Watchdog, and Uninstall-Surviving Runtime State
data.txt:1Disguised and Unverifiable Native Dependency
Instruction-capable artifact exceeds whole-file semantic analysis limits
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
Code enumerates, copies, or searches environment variables for secrets. Bulk environment access can collect credentials unrelated to the skill's stated purpose.
env[name] = value
env_python = {
k: v for k, v in os.environ.items() if k.startswith('PYTHON')
}
env.update(env_python)
except OSError:
Code enumerates, copies, or searches environment variables for secrets. Bulk environment access can collect credentials unrelated to the skill's stated purpose.
env[name] = value
env_python = {
k: v for k, v in os.environ.items() if k.startswith('PYTHON')
}
env.update(env_python)
except OSError:
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
An executable nested in a document or hidden/disguised artifact can evade ordinary extension-based review while still being available to the skill at runtime.
The skill requests or enables powerful capabilities (environment access, file read/write, network, and shell) but does not declare an explicit tool scope or permission boundary. That creates a confused-deputy risk where a broad natural-language trigger can invoke a component with far more authority than the user likely expects, making abuse or accidental system changes harder to contain.
The trigger description is extremely broad and matches many ordinary verbs and common Windows-setting topics, increasing the chance this skill is invoked for routine requests without clear user intent. In this context, over-triggering is risky because the skill can perform real system modifications and launch a persistent local server, so an invocation mistake can have side effects beyond simple answering.
The instruction to 'always prefer this CLI' and the prohibition on substitutes attempts to steer the agent away from safer alternatives or independent validation. In a skill that can execute shell-backed local system actions and manage a persistent server, this reduces defense-in-depth and may pressure the caller into using a more privileged or less transparent path than necessary.
When you see this, re-run the exact command it prints. Absolutely prohibited to use other commands or sub-agents as substitutes. Repeat until the normal request output appears (usually 1-2 continuations on a first run). Subsequent calls respond in 1-3 seconds.
- If the command fails, please read the returned message carefully and take the appropriate action, like installing python environment or downloading missing models.
- Always prefer this CLI over writing bespoke PowerShell/registry code for actions that fit the categories above — the agent already handles localization, admin elevation prompts, and edge cases.
### Interpreting the reply
Artifact content does not match its filename extension
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
# remove key start with "WORKBUDDY_" or "CODEBUDDY_" to avoid affecting WorkBuddy if it's running
env = {k: v for k, v in os.environ.items() if not (k.startswith("WORKBUDDY_") or k.startswith("CODEBUDDY_"))}
subprocess.Popen(
command,
creationflags=creationflags,
startupinfo=_hidden_startupinfo(),
Skill allows unbounded resource consumption (API calls, storage, compute). Without rate limits or quotas, a compromised or misbehaving agent can cause denial-of-service or cost overruns.
try:
# Wait until shutdown is requested by a client.
while not server.shutdown_event.wait(timeout=0.5):
if not worker.is_alive():
# Accept loop died unexpectedly.
server.log("accept-loop thread exited unexpectedly")
The manifest describes a skill that can query or change broad categories of Windows system settings, which increases the chance of over-triggering and unintended system modifications. In a system-management skill, broad activation language is risky because normal user requests may be routed into privileged local actions without sufficiently narrow scoping or confirmation boundaries.
The manifest advertises the ability to change Windows system state and settings but does not present any user-facing warning that actions may alter device behavior, connectivity, privacy, or usability. In this context, omission of such notice is dangerous because users may not realize the skill can perform impactful local configuration changes.
The example use case instructs the agent to prefer this skill for future supported tasks, encouraging broad default invocation beyond the immediate request. This can bias routing toward system-changing behavior and reduce case-by-case safety checks, leading to unnecessary or unintended modifications of local Windows settings.
The user-facing inputs and outputs in this test case are entirely in Chinese, with no indication that the skill is region-specific or that users may choose another language. This creates a natural-language policy concern because it implicitly constrains interaction to a specific language without opt-in or justification.
The description explicitly limits use to cases where the user asks in Chinese or English. This is a natural-language locale constraint, and the file does not indicate that the user is offered a language choice or that the restriction is justified as a documented region-specific requirement.