understand-image-video

Warn

Audited by Socket on Aug 24, 2026

2 alerts found:

AnomalySecurity
AnomalyLOW
SKILL.md

SUSPICIOUS. The skill’s capabilities largely match its stated vision-analysis purpose, and the main network flow to Agnes is disclosed. However, it reads local credential files and relies on unpinned uvx/PyPI/Hugging Face resolution, creating medium supply-chain and privacy risk. No evidence of overt malware, covert exfiltration endpoints, or fundamentally mismatched behavior was found.

Confidence: 86%Severity: 62%
SecurityMEDIUM
scripts/agnes-understand.py

The code is primarily an API client for multimodal understanding, but it contains a high-risk capability: when an API key entry in local config begins with '!', the script executes that string via subprocess.run(..., shell=True) to derive the API key. This constitutes arbitrary command execution from configuration content and could enable sabotage or persistence if config files are tampered with. Additionally, the tool exfiltrates local media contents by base64-embedding them into requests to a fixed remote API (privacy/security risk if misused).

Confidence: 78%Severity: 70%
Audit Metadata
Analyzed At
Aug 24, 2026, 12:13 AM
Package URL
pkg:socket/skills-sh/lalalic%2Fbullx%2Funderstand-image-video%2F@84a14ffddcfbe72d171d9be7e906acc8e238219266de27ee04eadb47123fa9ab