Back to skill

Security audit

lazy

Security checks for vulnerabilities and agentic risk

Overview

This is a coherent desktop automation skill, but it needs Review because it can silently expose screen or clipboard data and lets optional LLM output drive live desktop actions.

Install only if you are comfortable granting the skill control over your live desktop. Use failsafe and confirmation modes, avoid using it around passwords, private documents, chats, or financial/account pages, and do not connect an LLM client unless you accept that screenshots and task text may be shared with that client. Prefer running it in a disposable or low-privilege desktop session.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (5)

T09 · Insecure Skill Coding Practices

Error
Location
ai_agent.py:287
Finding

Unapproved Full-Screen Disclosure to an Arbitrary LLM Client

Content
View full analysis
List[Dict]: """Use LLM for intelligent task planning.""" # Capture current state screenshot = self.dc.screenshot() if screenshot: # Convert to base64 for LLM buffered = io.BytesIO() screenshot.save(buffered, format="PNG") img_b64 = base64.b64encode(buffered.getvalue()).decode() prompt = f"""Given this desktop screenshot, create a plan to: {task} Return a JSON list of steps, each with: - type: action type - description: what to do - [other relevant parameters] Available actions: launch, click, type, hotkey, wait, screenshot""" try: response = self.llm.generate(prompt, images=[img_b64]) ``` ### Technical Analysis When local pattern matching cannot plan a task and an LLM client is configured, the agent automatically captures the entire desktop. It serializes the screenshot as PNG, Base64-encodes it, and passes the encoded image to the caller-provided `LLMClient`. Base64 is a transport encoding and provides no confidentiality or access control. The `LLMClient` protocol does not constrain the implementation, endpoint, retention policy, or geographic destination. The code also lacks: - Explicit consent before transmitting the image - Screen-region restriction - Sensitive-window detection - Secret or personal-data redaction - Endpoint allowlisting - Data-classification checks - A local-only planning fallback Desktop screenshots can expose passwords, authentication tokens, private messages, customer records, source code, financial data, or other applications unrelated to the requested task. Sending the full desktop exceeds the minimum privilege required for many planning operations. The proj ...[truncated 1305 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
ai_agent.py:297
Finding

Untrusted LLM Plans Directly Control Privileged Desktop Actions

Content
View full analysis
StepResult: """Execute a single step and return result.""" step_type = step.get("type", "unknown") description = step.get("description", step_type) start_time = time.time() try: # Screenshot before action if enabled if self.screenshot_before_each_step: step["screenshot_before"] = self.dc.screenshot() # Execute based on type if step_type == "launch": self._do_launch(step["app"]) elif step_type == "click": x, y = step.get("x"), step.get("y") self.dc.click(x, y) elif step_type == "type": self.dc.type_text(step["text"], wpm=step.get("wpm", 100)) elif step_type == "hotkey": self.dc.hotkey(*step["keys"]) elif step_type == "press": self.dc.press(step["key"]) elif step_type == "wait": time.sleep(step["duration"]) elif step_type == "activate": self.dc.activate_window(step["title"]) elif step_type == "screenshot": self.dc.screenshot_to_file(step.get("filename", "screenshot.png")) elif step_type == "draw": self._do_draw(step["shape"]) elif step_type == "analyze": pass # Analysis step ``` ```python def _do_launch(self, app: str) ...[truncated 2957 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
desktop_control.py:309
Finding

Approval Mode Does Not Mediate All Sensitive Reads and Desktop Actions

Content
View full analysis
None: self._check_failsafe() if x is not None and y is not None: self.move_mouse(x, y, duration=0) import pyautogui if horizontal: pyautogui.hscroll(amount) else: pyautogui.scroll(amount) ``` Key hold and release operations are also unmediated: ```python def key_down(self, key: str) -> None: """Hold a key down.""" self._check_failsafe() import pyautogui pyautogui.keyDown(key) if self._log_debug: logger.debug(f"Key down: {key}") def key_up(self, key: str) -> None: """Release a held key.""" import pyautogui pyautogui.keyUp(key) if self._log_debug: logger.debug(f"Key up: {key}") @contextmanager def hold_keys(self, *keys: str): import pyautogui for key in keys: pyautogui.keyDown(key) try: yield self finally: for key in reversed(keys): pyautogui.keyUp(key) ``` Screenshot capture bypasses approval: ```python @retry_on_error(max_retries=2, delay=0.1) def screenshot( self, region: Optional[Tuple[int, int, int, int]] = None, filename: Optional[str] = None ) -> Optional['Image.Image']: import pyautogui img = pyautogui.screenshot(region=region) if filename: img.save(filename) logger.info(f"Screenshot saved: {filename}") return None ``` Window activation bypasses approval: ```python def activate_window(self, title: str, partial: bool = True) -> bool: try: gw = _get_pygetwindow() if partial: windows = gw.getWindowsWithTitle(title) ...[truncated 2597 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
desktop_control.py:378
Finding

Sensitive Typed and Clipboard Content Is Recorded in Plaintext Logs

Content
View full analysis
50 else text if self._check_approval(f"type: '{display_text}'"): import pyautogui pyautogui.write(text, interval=interval) logger.info(f"Typed '{display_text}' (interval={interval:.3f}s)") ``` ```python def copy_to_clipboard(self, text: str) -> bool: """Copy text to clipboard.""" try: pc = _get_pyperclip() pc.copy(text) logger.debug(f"Copied to clipboard: {text[:50]}...") return True ``` ### Technical Analysis The controller logs up to the first 50 characters of text typed through `type_text()` at `INFO` level and clipboard content at `DEBUG` level. The same text is also included in the approval prompt. Desktop automation commonly handles passwords, authentication tokens, email addresses, private messages, form data, and customer records. The project documentation itself demonstrates password entry, making secret-bearing input a foreseeable use case. Truncation is not redaction. Passwords and tokens are often shorter than 50 characters, while the leading portion of a longer secret can still enable identification, correlation, or partial compromise. Logs may be written to files, collected centrally, included in support bundles, or visible to other processes and users. ### Attack Path 1. A user uses `type_text()` to enter a password, token, personal record, or confidential message, or uses `copy_to_clipboard()` with sensitive content. 2. The controller constructs a log message containing the first 50 characters. 3. The active logging configuration writes or forwards the message. 4. An attacker, administrator, monitoring service, support recipient, or lower-trust process obtains access to the logs. 5. The sensitive content is recovered directly from the log e ...[truncated 534 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
requirements.txt:1
Finding

Unbounded Dependency Versions Create Non-Reproducible Supply-Chain Exposure

Content
View full analysis
=0.9.54 Pillow>=10.0.0 opencv-python>=4.8.0 pygetwindow>=0.0.9 pyperclip>=1.8.2 pyscreeze>=0.1.29 ``` ### Technical Analysis Every dependency is specified using only a minimum version. A future installation may therefore resolve to package releases that did not exist and were not reviewed when this Skill was published. These dependencies interact with high-value local resources, including the screen, keyboard, mouse, clipboard, image data, windows, and filesystem. Installation and import of a compromised future release or transitive dependency can execute code with the installing user's privileges. No malicious or typosquatted dependency was confirmed in the reviewed project. The vulnerability is the absence of reproducible version and integrity controls, not evidence that the named packages are currently malicious. ### Attack Path 1. A future upstream package release or transitive dependency is compromised, maliciously replaced, or introduces a severe vulnerability. 2. A user executes `pip install -r requirements.txt`. 3. The package resolver selects the affected version because the constraints have no upper bounds or exact pins. 4. Package installation hooks or imported package code execute in the user's environment. 5. The compromised dependency gains the privileges and data access available to the installation or Skill process. ### Impact Assessment A compromised dependency could obtain code execution with the privileges of the user installing or running the Skill. Given the nature of the dependencies, it could potentially access screen contents, clipboard data, local files, keyboard and mouse automation, and application windows. The exact scope depends on the affected package and environment. The reviewed files do not demonstrate an existing depende ...[truncated 116 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (32)

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The published description understates the breadth of the automation surface: beyond basic desktop control, it supports clipboard operations, window management, active-window detection, template matching, dialogs, and key polling. In an agent setting, understated capabilities are dangerous because reviewers or users may authorize a skill without understanding that it can read sensitive context and manipulate arbitrary applications.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The published description understates the breadth of the automation surface: beyond basic desktop control, it supports clipboard operations, window management, active-window detection, template matching, dialogs, and key polling. In an agent setting, understated capabilities are dangerous because reviewers or users may authorize a skill without understanding that it can read sensitive context and manipulate arbitrary applications.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

When LLM planning is enabled, the code captures a live desktop screenshot, base64-encodes it, and sends it to the configured LLM service together with the task prompt. Desktop screenshots may contain secrets, personal data, messages, tokens, or other sensitive on-screen content, and this external transmission is not disclosed in the skill description.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The agent sends a desktop screenshot to the LLM automatically during planning without any explicit warning or confirmation from the user. Because screenshots can expose highly sensitive information from unrelated applications, background windows, or notifications, this creates a silent privacy and data-exfiltration risk.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill is described as accepting broad natural-language instructions and autonomously determining actions, but it does not define clear scope boundaries or trigger constraints. In a desktop-control context, this can cause the agent to interpret unsafe or overly broad requests and perform unintended actions on the user's system.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The guide describes autonomous screenshots, screen observation, and file-affecting behavior without prominent warnings about privacy, sensitive data capture, or consent. In a desktop automation skill, screenshots and OCR can expose credentials, personal documents, messages, or regulated data, especially when actions occur autonomously.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The guide states that the agent can currently analyze screenshots, detect UI elements, read text via OCR, and identify objects, and even shows _analyze_screen() output as if implemented. However, the later 'Future Enhancements' section lists computer vision, OCR, and LLM integration as planned features, which directly contradicts the earlier documentation about present capabilities.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The documented use cases include form filling, social media posting, and copying data between documents and apps, all of which can affect personal data, external accounts, or public-facing content. Without warnings and approval gates, the agent could submit sensitive information, post unintentionally, or mishandle third-party account actions under broad natural-language prompts.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The documentation presents a 'failsafe=False' mode as a simple configuration option without clearly warning that it disables an important control against runaway or unsafe autonomous interaction. In a desktop agent that can click, type, launch apps, and operate on visible content, removing failsafes materially increases the risk of uncontrolled destructive or privacy-impacting actions.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The document advertises screenshots, pixel inspection, window enumeration, clipboard reads, and window activation without a clear privacy warning that these features may expose sensitive on-screen content, application titles, or clipboard secrets. In a desktop-control skill, these capabilities materially increase the risk of collecting or mishandling credentials, personal data, or confidential business information.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The quick reference normalizes high-impact desktop actions like typing, clicking, hotkeys, and form submission, but the safety section is brief and appears late in the document. In an automation skill, this can lead users or downstream agents to trigger unintended UI actions that modify files, submit forms, launch programs, or change system state without adequate upfront warning or gating.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This skill can generate real mouse and keyboard input, switch windows, and trigger actions in live applications, which can modify files, send messages, approve dialogs, or disrupt system state. Without an upfront warning, users may treat examples as harmless when they actually perform privileged real-world actions on the desktop.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

Screenshot capture, pixel inspection, image matching, and clipboard access can expose secrets visible on screen or stored in the clipboard, including credentials, personal data, and confidential documents. Documenting these features without a privacy warning normalizes sensitive collection and increases the chance of accidental data exposure by agent workflows.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The agent can launch applications based on natural-language task input, and its built-in app knowledge explicitly includes cmd, while the generic launch path can invoke arbitrary executables through the Run dialog. In a desktop automation skill, this meaningfully expands capability into arbitrary program execution, which can be abused to open shells or high-risk tools if untrusted or overly broad task input is accepted.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The task description is included in the prompt sent to the configured LLM without a specific disclosure to the user. Task text may contain sensitive instructions, filenames, account details, or business context, so sending it off-box can leak user data even when no screenshot is present.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The demo reads the current clipboard, overwrites it with new content, and may restore it afterward, but it does so without explicit user confirmation at the moment of access. Clipboard contents often contain sensitive data such as passwords, tokens, or personal information, so silent access and modification creates a privacy and data exposure risk even in a demo context.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The screenshot function can capture all or part of the user’s screen and optionally save it to disk, yet it does not require approval by default or present any user-facing warning. Screenshots can expose emails, chats, financial data, passwords visible on screen, or other sensitive application contents, and saving to file expands persistence and exfiltration risk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The code enumerates open windows, inspects the active window, and can activate windows, but the skill description does not disclose these surveillance and control capabilities. Window titles can reveal sensitive user activity, documents, websites, or applications, and activation can redirect user focus or facilitate further automated actions.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill description advertises mouse, keyboard, and screen control but does not disclose clipboard access, while the code can both read and write clipboard contents. Clipboard data often contains passwords, tokens, personal data, or copied secrets, so omitting this capability reduces informed consent and increases the risk of covert data access.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Clipboard copy and paste operations interact with highly sensitive user data but are exposed without disclosure or approval safeguards by default. Reading clipboard contents is particularly risky because it can silently collect whatever the user most recently copied, including secrets unrelated to the task.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Clipboard reading is broader than necessary for basic desktop automation and enables collection of arbitrary user-copied data without strong justification in the stated purpose. Because clipboard contents commonly include credentials, API keys, addresses, or confidential text, this materially increases privacy and data-exfiltration risk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Low
Category
Not specified by scanner
Confidence
86% confidence
Finding

Window management and enumeration reveal application context and enable focus-stealing or interaction with unintended windows. If these capabilities are not declared up front, users and reviewers may not realize the skill can inspect window titles and direct input into arbitrary applications.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

Clipboard read/write is a sensitive data-access capability because clipboard contents often contain passwords, tokens, private messages, or copied documents. Omitting it from the manifest-level description weakens informed consent and can cause downstream tooling to misclassify the skill as less privacy-invasive than it is.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
86% confidence
Finding

The demo captures screenshots and writes image files to disk without an explicit advance warning that files will be created containing screen contents. Screenshots can capture sensitive on-screen information, and saving them to disk increases persistence and the chance of later disclosure.

Content

No source excerpt is available for this finding.

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
96% confidence
Finding

The dependency is specified with only a minimum version, which allows newer, unreviewed releases to be installed. This weakens build reproducibility and can introduce vulnerable or malicious upstream changes without notice.

Content

Scanner excerpt · requirements.txt (reported line 1)May include surrounding context.

text
pyautogui>=0.9.54
Pillow>=10.0.0
opencv-python>=4.8.0
pygetwindow>=0.0.9