ClawHub
Back to skill

Security audit

Perf Test Flagos

Security checks across malware telemetry and agentic risk

Overview

This skill does what it says, but it runs high-impact container commands and enables remote model code execution by default without a clear opt-in warning.

Install only if you intend to let the agent control the specified benchmark container. Use trusted, pinned model/tokenizer sources, avoid --trust-remote-code for untrusted repositories, run in an isolated container without sensitive mounts, and replace the broad pkill stop step with a tracked process or dedicated container before shared use.

SkillSpector

By NVIDIA
Vulnerability Patterns
  • 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
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (2)

Missing User Warnings

Low
Confidence
81% confidence
Finding
The skill copies benchmark scripts into the container and writes benchmark outputs to `/data/results/perf` without an explicit warning that it will modify container state. While this is aligned with the benchmarking purpose and not inherently malicious, silent writes can surprise users, overwrite prior results, or affect reproducibility in shared or sensitive containers.

Missing User Warnings

Medium
Confidence
98% confidence
Finding
The script always passes --trust-remote-code to vllm bench serve, which can cause execution of model/tokenizer repository code during benchmark setup. Because this skill is designed to benchmark arbitrary served models and tokenizers, enabling remote code by default materially increases the risk of arbitrary code execution inside the container without any explicit opt-in or warning.

VirusTotal

VirusTotal findings are pending for this skill version.

View on VirusTotal

Static analysis

No suspicious patterns detected.