SkillSpector Security Scan Results

[lgdlair]

SkillSpector Security Scan Results

I ran NVIDIA SkillSpector v2.2.3 on ai-image-prompts-skill and found 33 security issues (score: 100/100 CRITICAL).

Summary

Severity	Count	Category
HIGH	~20	P6 - Direct Prompt Extraction (community prompt JSONs contain "ignore previous instructions" patterns)
HIGH	4	PE3 - Credential Access (SSH key paths, API keys in community prompts)
HIGH	1	RA1 - Self-Modification (generate-references.ts)
HIGH	1	SC1 - Supply Chain (unpinned dependencies in package.json)
MEDIUM	4	YR4 - Network reconnaissance references in prompts (nmap, scanning tools)
MEDIUM	1	TM1 - Tool Parameter Abuse (exec: rm, exec: curl without validation)
MEDIUM	1	LP3 - No declared permissions
LOW	1	EA3 - Scope Creep

Root Cause

The vast majority of HIGH-severity issues come from user-submitted prompts in references/*.json. These JSON files are community-contributed and contain:

• SSH private key paths (/Users/xxx/.ssh/id_rsa)
• Prompt injection attempts ("ignore previous instructions")
• Network scanning tool references (nmap)
• Hidden instructions

Since references/*.json are auto-generated daily from the YouMind platform, they reflect what community users upload without sanitization.

Suggested Fixes

1. Filter community prompts before generating reference JSONs — scan for credential patterns, injection attempts, and malicious content at submission time.
2. Pin dependencies in package.json to avoid supply chain risks.
3. Add permissions field to SKILL.md.
4. Wrap exec commands with input validation or use safer APIs.

---

Generated by SkillSpector v2.2.3