Back to skill

Security audit

Xia Desktop Agent

Security checks across malware telemetry and agentic risk

Overview

This desktop automation skill is powerful and mostly transparent about its goal, but it handles remote-access credentials, screenshots, app launching, and WeChat actions with too little scoping or confirmation.

Install only if you fully trust the publisher and intend to give the skill broad control over your Windows desktop. Before use, require explicit confirmation for ToDesk credential sharing, WeChat sends, file transfers, app launches, and file writes; avoid using it while sensitive windows are visible, and clear or restrict stored screenshots.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (19)

subprocess module call

Medium
Category
Dangerous Code Execution
Content
"code": "code",
            }
            cmd = app_map.get(app_name.lower(), app_name)
            subprocess.Popen(cmd, shell=True)
            logger.info(f"启动应用: {app_name} -> {cmd}")
            return True
        except Exception as e:
Confidence
98% confidence
Finding
subprocess.Popen(cmd, shell=True)

Lp3

Medium
Category
MCP Least Privilege
Confidence
78% confidence
Finding
The skill exposes capabilities consistent with network and shell-adjacent execution but does not declare permissions or clearly communicate those powers. In a desktop automation skill, undeclared execution and communication surfaces reduce transparency and make abuse or unexpected data flow harder to detect and constrain.

Tp4

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding
The documented purpose says desktop automation, but the described behavior extends into local LLM prompt transmission, window enumeration, screen-based element discovery, and direct Windows message injection. That mismatch matters because users may consent to simple GUI automation without realizing the skill can inspect broader desktop context and inject input into other applications.

Context-Inappropriate Capability

Medium
Confidence
79% confidence
Finding
Enumerating all visible window titles and geometry exposes sensitive contextual information such as open apps, document names, chats, browser tabs, and screen layout. While useful for automation, it also provides reconnaissance beyond a narrowly scoped user action and can aid targeting of subsequent clicks, keystrokes, or data theft.

Intent-Code Divergence

Medium
Confidence
99% confidence
Finding
The module is explicitly described as preventing dangerous operations, but both check_task() and check_plan() only log matches and still return success for dangerous content. In a desktop automation skill that can click, type, open apps, send WeChat messages, and establish remote connections, this creates a misleading safety boundary and allows destructive or unauthorized OS actions to proceed unchecked.

Vague Triggers

Medium
Confidence
88% confidence
Finding
Broad activation phrases like '远程连接', 'ToDesk', or '连接你' can be triggered in ordinary conversation and map directly to remote-control behavior. For a skill that can launch remote-access software and extract credentials, loose invocation language increases the risk of accidental or socially engineered activation.

Vague Triggers

Medium
Confidence
85% confidence
Finding
The description advertises use for essentially any Windows GUI task, which is overly broad for a skill with powerful automation primitives. Such wide invocation scope increases the chance that unrelated user requests are routed into a high-privilege desktop controller without clear boundaries.

Missing User Warnings

High
Confidence
97% confidence
Finding
The skill explicitly describes sending ToDesk credentials to the user but provides no sensitivity warning, identity verification, or consent guardrail. Remote-access credentials are highly sensitive, and automating their extraction and disclosure from the desktop materially increases the risk of unauthorized access.

Missing User Warnings

Medium
Confidence
82% confidence
Finding
The natural-language workflow promises autonomous arbitrary desktop control and file-saving without prominent user warning about side effects, data exposure, or application focus issues. In this context, even benign tasks can manipulate other windows, overwrite files, or leak visible on-screen content through screenshots and automation.

Missing User Warnings

Medium
Confidence
86% confidence
Finding
Screenshots are written to a persistent directory on disk by default, which can retain sensitive desktop content such as messages, credentials, documents, and personal data. Because the file is saved without explicit user-facing consent or retention controls, it increases the risk of unintended disclosure and later access by other processes or users.

Missing User Warnings

Medium
Confidence
84% confidence
Finding
The fallback input path copies arbitrary text to the system clipboard, which can overwrite user clipboard contents and expose sensitive data to other applications or clipboard history mechanisms. In a desktop automation context, this side effect is non-obvious and may leak secrets or interfere with the user's session.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The skill allows arbitrary application launch and maps unknown inputs directly to a command string, then executes it through the shell. Without confirmation, allowlisting, or command validation, this permits execution of unintended programs or attacker-crafted commands, which is highly risky on a desktop-control agent.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The function can send outbound WeChat messages immediately after receiving parameters, with no in-function confirmation, preview, recipient verification, or policy gate. In a desktop-control skill, this creates a realistic risk of unauthorized messaging, social engineering, spam, or accidental transmission to the wrong contact.

Missing User Warnings

High
Confidence
98% confidence
Finding
This function automates file transmission from an arbitrary local path to a selected WeChat contact without any final consent check or validation of file sensitivity. In the context of a desktop automation agent, that materially increases the risk of silent exfiltration of confidential files, including documents outside the user's intended scope.

Missing User Warnings

High
Confidence
97% confidence
Finding
The workflow launches remote-access software, captures a screenshot of credential-bearing UI, and returns a hard-coded device code plus a path to the screenshot for later password recognition. In this skill's context, that enables collection and sharing of remote-access credentials, which could facilitate unauthorized persistent access to the user's machine if misused or mishandled.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The executor automatically captures screenshots before each step and again at the end, then stores their references in the execution log without any user-facing notice, consent flow, or data-minimization control. In a desktop-automation skill that can open apps, type text, interact with WeChat, and establish remote connections, those screenshots can expose sensitive messages, credentials, documents, or other on-screen data far beyond what is needed for task completion.

Missing User Warnings

Medium
Confidence
92% confidence
Finding
The function sends the full user task to a local LLM HTTP endpoint without any user-facing disclosure, consent, or data-minimization controls. In a desktop automation skill, user tasks may contain sensitive data such as message contents, credentials, internal documents, or remote-access instructions, so silent transmission increases privacy and compliance risk even if the endpoint is localhost.

Ssd 3

High
Confidence
98% confidence
Finding
The instructions tell the agent to extract a dynamically generated remote-access password from a screenshot and disclose it. This is effectively credential harvesting and transmission, and in a desktop-control skill the surrounding context makes it particularly dangerous because the same agent can also launch the remote-access tool and operate the machine.

Ssd 3

High
Confidence
98% confidence
Finding
The workflow directs the agent to read a device code and temporary password from an image and return them, which is direct extraction of sensitive authentication material. Because the skill is designed for desktop automation and remote connection setup, the context amplifies the risk from mere data exposure to potential full remote compromise.

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.