The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Piped execution of a remote script from an unknown source. The SKILL.md file instructs users to run curl -fsSL https://treg.to/install.sh | sh. This pattern is highly insecure as it downloads and executes arbitrary code directly from an untrusted domain without any integrity verification.
[EXTERNAL_DOWNLOADS]: Interaction with multiple non-whitelisted third-party file hosting services. The scripts/host_file.py utility attempts to upload and verify assets on uguu.se, catbox.moe, and litterbox.catbox.moe. These services are not standard for asset handling and are often associated with transient or untrusted content.
[COMMAND_EXECUTION]: Extensive use of shell command execution via subprocess calls. Multiple scripts (caption_burn.py, seedance_treg.py, and host_file.py) execute external tools including ffmpeg, ffprobe, curl, and the treg CLI to process video and handle network operations.
[INDIRECT_PROMPT_INJECTION]: The skill maintains a vulnerability surface for indirect prompt injection attacks.
Ingestion points: The skill processes user-provided script text, character images, and voice reference clips as inputs for video generation.
Boundary markers: There are no delimiters or protective instructions used to isolate untrusted user content from the internal prompt logic.
Capability inventory: The skill possesses capabilities to execute shell commands, perform network uploads/downloads to external hosts, and write files to the local system.
Sanitization: No validation, escaping, or sanitization is performed on user-supplied data before it is interpolated into prompts or processed by local execution scripts.

ugc-talking-head-video