Play Any Game - AI游戏伴侣助手

• Outcome: Review
• Latest audit: May 29, 2026 · 9:32 PM UTC
• Version: 1.4.1

Overview

This game-helper skill is not clearly malicious, but it has broad screen-viewing and computer-control powers with weak limits, local secret storage, and automation guidance that users should review carefully.

Install only if you are comfortable giving the skill access to screenshots, saved local images, a DashScope API key, and mouse/keyboard control of your Windows desktop. Prefer dry-run or advisory use, restrict it to a visible game window, avoid background/global clicks, and do not use it for unattended farming or actions that could affect purchases, accounts, files, or non-game windows.

SkillSpector

By NVIDIA

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep
• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• Behavioral AST: exec() Call, eval() Call, Dynamic Import
• MCP Tool Poisoning: Hidden Instructions, Unicode Deception, Parameter Description Injection

Findings (33)

eval() call detected

High

• Category: Dangerous Code Execution

Content

if isinstance(expr, str):
        expr = expr.replace('width', 'base').replace('height', 'base')
        try:
            return int(eval(expr, {'base': base_value}))
        except:
            return 0

• Confidence: 98% confidence
• Finding: return int(eval(expr, {'base': base_value}))

Description-Behavior Mismatch

Medium

• Confidence: 98% confidence
• Finding: The skill content defines a broad general-purpose assistant with capabilities far beyond a game companion, including technical help, writing, automation, and web tasks. In the context of a game-helper skill, this scope expansion increases the chance the agent will accept unrelated high-risk requests and bypass expected least-privilege boundaries.

Context-Inappropriate Capability

High

• Confidence: 99% confidence
• Finding: Advertising the ability to operate computer software and files is dangerous because it grants or normalizes system-level actions unrelated to helping with games. In a game-companion context, this can lead to unauthorized file modification, destructive actions, or misuse of host applications under the guise of assistance.

Context-Inappropriate Capability

Medium

• Confidence: 97% confidence
• Finding: Broad automation, web search, and code writing/debugging materially exceed the declared purpose of a game-helper skill. This overbroad authority enlarges the attack surface by enabling unrelated external communication or code-related assistance that users would not reasonably expect from a narrowly scoped gaming companion.

Description-Behavior Mismatch

Medium

• Confidence: 96% confidence
• Finding: The example of organizing desktop files demonstrates a concrete off-scope action that can modify user data, reinforcing that the skill may act as a general desktop agent rather than a game assistant. Examples strongly shape agent behavior, so including this workflow makes misuse more likely in real interactions.

Description-Behavior Mismatch

Medium

• Confidence: 88% confidence
• Finding: The skill is described as a game-assistance helper, but it also exposes configuration and agent-management operations that are outside that stated scope. Scope creep matters for security because it gives the caller extra capabilities, including credential handling and agent selection, that are not necessary for helping a user get past a game screen and enlarge the attack surface.

Description-Behavior Mismatch

Medium

• Confidence: 92% confidence
• Finding: The code can enumerate visible windows and activate arbitrary windows system-wide, not just the intended game window. In the context of a skill that also performs clicks, key presses, screenshots, and text-driven automation, this broad targeting capability can be used to interact with unrelated applications and increases the risk of unintended or unauthorized actions outside the game.

Context-Inappropriate Capability

Medium

• Confidence: 87% confidence
• Finding: The skill reads, stores, and displays management paths for external AI service credentials even though credential administration is not inherent to a game-helper skill. Handling API keys inside a broadly accessible automation tool increases the chance of secret exposure or misuse, especially when combined with external screenshot-upload features.

Context-Inappropriate Capability

High

• Confidence: 97% confidence
• Finding: If no window is identified, the code falls back to a global screen click via pyautogui, allowing unconstrained interaction anywhere on the desktop. In a skill that accepts natural-language targeting and screenshots, this can misfire into browsers, terminals, password prompts, or other sensitive applications and cause destructive or privacy-impacting actions.

Description-Behavior Mismatch

Medium

• Confidence: 94% confidence
• Finding: The document explicitly defines a repeatable stamina-farming loop: check resources, start challenge, wait for combat, optionally enable auto-battle, collect rewards, and repeat until stamina is exhausted. That goes beyond contextual gameplay assistance and enables unattended grinding behavior, which can violate platform/game terms and expand the skill into automation rather than user-directed help.

Description-Behavior Mismatch

Medium

• Confidence: 92% confidence
• Finding: This tower-mode section describes a multi-floor automation cycle that repeatedly starts challenges, detects outcomes, collects rewards, advances floors, and continues until a target completion condition is met. In the context of a skill advertised as a helper rather than a full auto-grinder, this materially increases the capability for repeated autonomous progression.

Description-Behavior Mismatch

High

• Confidence: 99% confidence
• Finding: The 'risk detection' guidance advises adding random delays and varying click intervals, which is characteristic of evasion techniques meant to reduce detection of automated behavior. Because this skill operates in a gaming automation context, such anti-detection guidance makes the capability substantially more dangerous by facilitating covert botting and undermining the manifest's claim that it is not for full automation.

Description-Behavior Mismatch

Medium

• Confidence: 97% confidence
• Finding: The code transmits the user's screenshot and instruction to DashScope via `client.chat.completions.create`, including the full image as a base64 data URL and the raw text prompt. For a game-helper skill, screenshots may still contain usernames, chats, notifications, or other desktop content, so undisclosed remote transmission creates a real privacy and data-handling risk even if the implementation is functionally intended.

Context-Inappropriate Capability

Medium

• Confidence: 90% confidence
• Finding: The code enumerates all visible desktop windows and exposes their titles/handles, which goes beyond the stated purpose of assisting with a specific user-invoked game window. In the context of an agent skill that can help operate UI, this broad discovery capability increases privacy and abuse risk because it can reveal unrelated applications and be used to target non-game windows.

Missing User Warnings

Medium

• Confidence: 90% confidence
• Finding: The README advertises screenshot analysis and direct clicking on the user's game UI without clearly stating that the skill can affect the live game session or that actions should remain user-confirmed. In a skill that interacts with a real desktop/game window, omission of consent, scope, and side-effect warnings can mislead users into granting automation they do not fully understand, increasing risk of unintended clicks, purchases, destructive in-game actions, or disclosure of on-screen information.

Missing User Warnings

Medium

• Confidence: 92% confidence
• Finding: The feature list normalizes continuous screenshot capture and automated UI interaction but does not disclose privacy or integrity implications. Because this skill's core context is observing the screen and acting on the interface, missing warnings materially increase the chance that users expose chats, account details, or trigger unintended in-game/system actions without informed consent.

Vague Triggers

Medium

• Confidence: 89% confidence
• Finding: The activation phrases are very broad and overlap with normal conversation such as asking for help or saying something is hard to operate. That can cause accidental invocation of a skill with screenshoting and UI-control capabilities in contexts unrelated to gaming, unexpectedly triggering sensitive desktop interaction.

Missing User Warnings

Medium

• Confidence: 93% confidence
• Finding: The markdown promotes a loop of screenshot analysis and automatic UI operation but does not clearly warn that screenshots may contain sensitive information and may be sent to an external cloud provider, nor that input automation can affect the wider desktop if targeting is wrong. In this context, the combination of screen capture, external transmission, and desktop control materially raises privacy and system-safety risk.

Missing User Warnings

Medium

• Confidence: 88% confidence
• Finding: Network information retrieval without a privacy or transmission warning can expose user prompts, game context, or other data to external services without clear notice. While lower impact than file manipulation, it still creates an avoidable privacy risk, especially when such web access is unnecessary for a narrowly defined game assistant.

Missing User Warnings

Low

• Confidence: 88% confidence
• Finding: Network information retrieval without a privacy or transmission warning can expose user prompts, game context, or other data to external services without clear notice. While lower impact than file manipulation, it still creates an avoidable privacy risk, especially when such web access is unnecessary for a narrowly defined game assistant.

Missing User Warnings

Medium

• Confidence: 97% confidence
• Finding: The example response commits to reorganizing desktop files immediately and without confirmation, caution, or a description of what will change. This normalizes direct modification of user data and could lead to harmful or irreversible actions if the model follows the pattern literally.

Missing User Warnings

Medium

• Confidence: 90% confidence
• Finding: The skill explicitly documents screenshot capture and both foreground and background input injection into a live game window, but it does not include any user-facing warning, confirmation, or safety guidance about privacy, accidental clicks, or unintended actions. In this context, background control is especially risky because it can trigger in-game purchases, account-affecting actions, or capture sensitive on-screen information without the user's active attention.

Missing User Warnings

Medium

• Confidence: 90% confidence
• Finding: The skill explicitly advertises real-time screenshot analysis of the user's game window but does not disclose what may be captured, how long screenshots are retained, or whether sensitive overlays, chats, notifications, or account information could be exposed. Even though this is framed as gameplay assistance, continuous visual access to a user's screen creates a real privacy risk if users are not clearly warned and given consent controls.

Missing User Warnings

Medium

• Confidence: 92% confidence
• Finding: The skill supports mouse and keyboard automation in a live game context without prominently warning about unintended clicks, destructive in-game actions, policy violations, or account consequences. Because the capability can directly affect game state, purchases, rewards, movement, or interactions, lack of guardrails and informed consent makes this a genuine safety issue.

Missing User Warnings

Medium

• Confidence: 94% confidence
• Finding: The skill explicitly advertises real-time screenshot viewing of the user's game window, but it does not clearly warn that screenshots may capture sensitive information such as account identifiers, chat messages, overlays, notifications, or other on-screen data. Because the capability is framed as continuous observation and even proactive monitoring, the context increases privacy risk beyond a one-off user-initiated screenshot.

VirusTotal

67/67 vendors flagged this skill as clean.

Static analysis

Detected: suspicious.dynamic_code_execution

Dynamic code execution detected.

Critical

• Code: suspicious.dynamic_code_execution
• Location: scripts/recognition.py:54