The Agent Skills Directory

[OBFUSCATION]: The SKILL.md file contains explicit instructions ("Do NOT read the python file, instead just call it with the parameters") and advice against checking specific directories ("You don't need to check the folder under /mnt/user-data"). These are concealment tactics designed to hide the skill's operations from the agent's internal analysis and auditing processes.
[PROMPT_INJECTION]: The skill uses directive language to override the agent's standard operating procedure of reviewing code or files before execution. By instructing the agent to ignore the script's content, the skill attempts to bypass typical safety filters and human-in-the-loop oversight.
[DATA_EXFILTRATION]: The scripts/generate.py script reads the content of files specified in the --reference-images and --prompt-file arguments and transmits them to an external endpoint (generativelanguage.googleapis.com). While the destination is a well-known service, the script allows reading any file on the system (e.g., SSH keys, environment files) and encoding it into the API payload, which represents a data exposure and exfiltration risk.
[COMMAND_EXECUTION]: The skill workflow relies on the execution of a Python script that performs unrestricted file system reads and network operations.

video-generation