Security audit

Household Airflow Comfort Map

Security checks across malware telemetry and agentic risk

Overview

This is a document-only home comfort skill, but its examples undermine its own safety limits by steering the agent into HVAC, window-safety, product, and health-adjacent advice.

Review before installing. The package does not appear to contain malware or code-execution risk, but its prompt examples may cause an assistant to give overconfident home ventilation, HVAC, product, or health-adjacent recommendations. It should be revised so examples stay limited to observation-based comfort routines and consistently defer equipment, symptoms, hazardous air, and safety concerns to qualified guidance.

SkillSpector

By NVIDIA

Vulnerability Patterns

MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (6)

Description-Behavior Mismatch

Medium

Confidence: 88% confidence
Finding: The README example expands a comfort-mapping skill into HVAC-style diagnosis and equipment advice by attributing room temperature issues to duct distance/obstruction and recommending booster fans. That conflicts with the stated safety boundary and can cause the agent to give users overconfident building-system guidance outside its intended scope.

Description-Behavior Mismatch

High

Confidence: 95% confidence
Finding: The README includes specific window-ventilation instructions such as leaving windows open while AC runs and prescribing opening widths, despite the manifest saying it avoids unsafe window use. This is dangerous because it normalizes potentially unsafe ventilation behavior without accounting for outdoor air quality, security, weather, building rules, or appliance-specific hazards.

Description-Behavior Mismatch

Medium

Confidence: 90% confidence
Finding: The example moves beyond comfort planning into health and air-quality guidance by discussing weakness, dizziness, stale air, fresh-air systems, and reducing CO2 concentration. In this context, that can mislead users into relying on non-medical, non-validated advice for symptoms that may indicate a health or indoor-environment hazard.

Description-Behavior Mismatch

Medium

Confidence: 94% confidence
Finding: The usage scenarios materially expand the skill from observational comfort planning into HVAC diagnosis, product recommendations, and energy-bill analysis. This is dangerous because downstream agents may follow the concrete examples over the stated safety boundary, causing the skill to produce advice outside its intended competence and potentially unsafe or misleading recommendations.

Intent-Code Divergence

High

Confidence: 98% confidence
Finding: The expected outputs directly contradict the documented safety boundary by prescribing prohibited actions, including HVAC causal analysis, device recommendations, specific fan-placement strategies, and air-quality claims. These examples are especially risky because examples in prompt skills often act as strong behavioral instructions, overriding earlier constraints and leading the model to generate unsafe, unauthorized, or medically adjacent guidance.

Natural-Language Policy Violations

Medium

Confidence: 85% confidence
Finding: Scenario 4 hardcodes a Chinese-language, China-specific response pattern without any stated locale gating or user preference requirement. While not directly a safety exploit, it can cause inappropriate region-specific advice, assumptions about products and housing norms, and inconsistent behavior that bypasses the skill's general safety framing.

VirusTotal

63/63 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.