comfyui

SUSPICIOUS. The skill is mostly aligned with its stated ComfyUI purpose and primarily talks to a local server, but it expands trust to third-party MCP/custom-node code, forwards optional API keys to those components, and can autonomously start processes and download models. This is a medium-risk operational skill rather than confirmed malware.

Overall security posture is dominated by a described network-exposed server-side execution feature: a custom aiohttp route (/radiance/terminal) that performs exec() into a persistent namespace using user-provided input when enabled via RADIANCE_ENABLE_TERMINAL. If this endpoint is reachable by untrusted parties, it constitutes a high-likelihood RCE/backdoor abuse scenario. Other imaging/VFX features (HDR/WebGL pipeline, model/LoRA utilities, sidecar handling, and export/QC) appear largely consistent with legitimate functionality, though the acceptance of external model artifacts can introduce supply-chain integrity risk. Recommendation: treat the terminal feature as a critical finding requiring immediate code review for authentication, network exposure controls, and safe execution/sandboxing, or remove the exec capability entirely.

No direct evidence of intentional malware (no eval/exec, subprocess, credential theft, or exfiltration). However, the module has meaningful security risks: it writes downloaded content to disk using a filename provided by the remote server without sanitization (path traversal/arbitrary overwrite risk), and it uses COMFY_HOST from the environment to control where it sends requests (SSRF-like targeting if the environment/endpoint is untrusted). These should be addressed by enforcing safe path handling (e.g., allowlisting/sanitizing filenames and validating final resolved paths under outdir) and restricting/validating the destination host/port (prefer localhost or allowlisted addresses, and use HTTPS where appropriate).