Back to skill

Security audit

观势

Security checks across malware telemetry and agentic risk

Overview

This appears to be a real strategy-analysis skill, but it needs review because it can persist sensitive analysis files and install additional skills from the network without strong scoping or provenance controls.

Review this before installing if you work with sensitive company strategy data. Avoid running the initializer with --yes unless you intentionally want it to create local skill dependencies, and treat the generated PPT outline as a persisted copy of potentially confidential analysis.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (10)

Lp3

Medium
Category
MCP Least Privilege
Confidence
84% confidence
Finding
The skill declares no permissions, yet its documented behavior includes file writing, network access, and shell-like installation flows. That creates a dangerous trust gap: a user or platform may treat it as a pure analysis skill while it can modify the local environment and fetch remote content.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding
This is a strong description-behavior mismatch: the skill presents itself as a strategy-analysis tool but also documents local filesystem changes, skill discovery, remote package download, extraction, and self-bootstrapping of subskills. Hidden operational behavior behind an analytical facade increases the chance of unauthorized code or content introduction into the environment.

Description-Behavior Mismatch

Medium
Confidence
76% confidence
Finding
Silent generation of a PPT outline file and invocation of another presentation skill expands the skill from analysis into persistent file creation and chained execution. Even if intended as convenience, hidden side effects can surprise users and create opportunities for unsafe downstream behavior through the delegated skill.

Intent-Code Divergence

Medium
Confidence
95% confidence
Finding
The file first states that when information is insufficient the agent must 'absolutely not' output analysis or advice, but later permits a '初步判断框架' as an appetizer before asking questions. That exception weakens the hard gate and creates a prompt-level path for the agent to provide substantive strategic guidance despite missing required facts, which can lead to unreliable or unsafe recommendations.

Description-Behavior Mismatch

Medium
Confidence
97% confidence
Finding
The initializer downloads ZIP content from a remote service and extracts it into the local skills directory, effectively installing code/content outside the manifest's stated strategic-analysis purpose. This creates a supply-chain risk: a compromised endpoint, tampered archive, or malicious dependency can persist locally and influence future agent behavior.

Missing User Warnings

Medium
Confidence
90% confidence
Finding
The skill explicitly states it will silently pre-generate a PPT outline and write it to disk without warning the user. Undisclosed file creation violates user expectations for an analysis skill and can be abused to leave artifacts, overwrite files in expected output paths, or trigger downstream workflows without informed consent.

Missing User Warnings

Medium
Confidence
94% confidence
Finding
The protocol explicitly requires silently writing `output/guanshi-ppt-outline.json` during the same skill invocation, before any user-facing confirmation is obtained. This creates an unauthorized side effect on the workspace and can surprise users, leak sensitive strategic content into persisted files, or trigger downstream automation based on artifacts the user did not knowingly approve.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
When `--yes` or force-like non-interactive paths are used, the script can proceed with network retrieval and local installation without a prominent user-facing warning at the moment of action. That weakens informed consent and makes unattended or scripted execution capable of introducing new local skills from the network unexpectedly.

Ssd 4

Medium
Confidence
96% confidence
Finding
Allowing a seemingly harmless three-sentence 'appetizer' before required gating is a classic gradual-bypass pattern: once the model is allowed to start framing the answer, it may continue into analysis, assumptions, or recommendations. In a strategy skill, even brief preliminary guidance can anchor decisions on incomplete data and effectively defeat the intended hard stop.

Unvalidated Output Injection

High
Category
Output Handling
Content
if entry.is_dir() and (entry / "SKILL.md").exists():
                    installed.add(entry.name)
    try:
        result = subprocess.run(["hermes", "skills", "list"], capture_output=True, text=True, timeout=10)
        for line in (result.stdout + result.stderr).split("\n"):
            parts = [p.strip() for p in line.split("│")]
            if len(parts) >= 2 and parts[1]:
Confidence
95% confidence
Finding
Model output is used without validation or sanitization. Unvalidated output injected into downstream contexts (SQL, shell, HTML) enables injection attacks and arbitrary code execution.

VirusTotal

59/59 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.