The Agent Skills Directory

[PROMPT_INJECTION]: The skill documentation includes a parameter disable_safety_filter which is set to true by default. This instruction explicitly directs the underlying model to bypass content safety guidelines.
[COMMAND_EXECUTION]: The skill uses the belt CLI within a Bash environment to run the pruna/p-video-avatar application. It constructs shell commands using input parameters like voice_script, video_prompt, and voice_prompt, which may contain untrusted data.
[EXTERNAL_DOWNLOADS]: The skill references an external installation guide hosted on GitHub (raw.githubusercontent.com/inference-sh/skills/) and prompts for external media URLs (images and audio) to be processed by the tool.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface.
Ingestion points: Untrusted data enters the agent context through the voice_script, video_prompt, and voice_prompt parameters in SKILL.md.
Boundary markers: There are no explicit delimiters or warnings to ignore instructions embedded within the processed text.
Capability inventory: The skill is configured to execute shell commands via the belt CLI tool using the Bash capability.
Sanitization: The skill lacks apparent sanitization or validation logic for the input strings before they are passed to the CLI tool.

p-video-avatar