The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shell utilities (grep and perl) to process files provided by the user, which presents a command injection risk. In references/humanize.md, the instructions define shell commands for file cleaning: grep -n ... "" and perl -CSD -pi -e ... "". Since the parameter is user-controlled through the /video-script humanize command, a maliciously crafted filename could lead to the execution of arbitrary shell commands on the host environment.
[DATA_EXFILTRATION]: The skill's humanize mode accepts a file path as an argument and uses perl -i (in-place edit) to modify its contents. This allows for the targeting of sensitive files (e.g., .env, .ssh/config, or credential files) if the user provides an arbitrary path, leading to unintended data modification or corruption.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content (pasted text or external files) for humanization and script rewriting, which is a common attack surface for indirect prompt injection. 1. Ingestion points: Untrusted data enters the agent context via the humanize and rewrite commands in SKILL.md. 2. Boundary markers: There are no specified delimiters or 'ignore embedded instructions' warnings around the data being processed. 3. Capability inventory: The skill has the ability to execute shell commands (grep, perl) and write files to the video-agent/scripts/ directory. 4. Sanitization: There is no logic present to detect or sanitize malicious instructions embedded within the text or files being humanized.

video-script