Security audit

S2-SWM Swarm Sync Protocol(群体同步与路权博弈引擎)

Security checks across malware telemetry and agentic risk

Overview

This skill matches its swarm-coordination purpose, but it claims PKI-backed security while the executable only checks a hard-coded signature before enabling sensitive robot coordination decisions.

Review carefully before installing, especially for any real robot, drone, vehicle, or safety-sensitive environment. Treat the current authentication as a demonstration placeholder, not production security; require real mutual authentication, signed messages, replay protection, encrypted channels, explicit authorization, and clear privacy controls before deployment.

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 (3)

Description-Behavior Mismatch

Medium

Confidence: 88% confidence
Finding: The skill metadata claims secure inter-agent interaction via cryptographic authentication, but the whitepaper instead specifies broad coordination behaviors such as federated perception, right-of-way arbitration, and physical collaboration. This scope mismatch is dangerous because operators may rely on the skill for security guarantees it does not actually provide, enabling unsafe trust decisions in a physical multi-agent environment.

Intent-Code Divergence

Medium

Confidence: 96% confidence
Finding: The document describes a 'handshake' as silent coordinate broadcasting and overlap detection, but provides no cryptographic authentication, freshness checks, or anti-spoofing protections. In a swarm or robotics context, an unauthenticated handshake can let a malicious or rogue agent impersonate peers, inject false presence data, and influence right-of-way or coordination decisions in the physical world.

Missing User Warnings

Medium

Confidence: 90% confidence
Finding: The protocol instructs agents to silently broadcast precise location and share sensor/tensor data P2P without notice, consent, minimization, or transmission safeguards. In this context, that creates both privacy and security exposure: sensitive physical location, environmental sensing, and inferred scene data could be intercepted, misused, or overshared to unauthorized agents, amplifying surveillance and attack surface.

VirusTotal

67/67 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.