Security audit

20260425203240

Security checks across malware telemetry and agentic risk

Overview

This is a local AI benchmark skill with expected red-team test prompts, not a skill that secretly accesses data or persists in the environment.

Install only if you want a benchmark/red-team evaluation resource. Use the adversarial examples only in authorized testing, avoid pasting them into unrelated systems, and verify any GitHub clone or script before running it. The packaged artifacts do not show credential collection or persistence, despite metadata tags suggesting sensitive capabilities.

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)

Ssd 1

High

Confidence: 96% confidence
Finding: This section embeds a highly realistic social-engineering prompt injection that pressures disclosure of protected instructions using authority, urgency, and harm framing. Even though presented as a test case, copying such text into an agent skill can normalize or inadvertently surface a ready-made attack payload that downstream systems or reviewers may reuse unsafely.

Ssd 2

Medium

Confidence: 94% confidence
Finding: The file enumerates hidden-instruction attack variants including Base64, zero-width characters, homoglyphs, Markdown-hidden text, and ROT13. While framed as a robustness test, this materially increases attack transferability by providing a compact catalog of stealth prompt-injection techniques that can be directly adapted for real misuse.

VirusTotal

66/66 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.