The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to download and execute a script directly from a remote URL (https://cli.inference.sh) using the curl -fsSL ... | sh pattern. This method of installation is highly insecure as it allows the remote server to execute arbitrary code on the user's system without any integrity checks or prior verification of the script's content.
[COMMAND_EXECUTION]: The skill relies on the execution of a third-party CLI tool (infsh) via the Bash tool. While the allowed-tools field correctly limits the scope to this specific command, the primary installation method for the tool itself is the high-risk remote execution mentioned above.
[EXTERNAL_DOWNLOADS]: The skill documentation suggests installing further extensions using npx skills add. This pulls code from remote repositories and executes it, introducing additional external dependencies into the agent's environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, such as audio and images from URLs, and generates text transcripts (e.g., via fast-whisper-large-v3). If these external sources contain hidden instructions, they could influence the agent's behavior in subsequent steps of the media pipeline, such as translation or further command execution.
Ingestion points: Audio, image, and video URLs provided to infsh app run commands; output from transcription models.
Boundary markers: None identified in the provided instructions.
Capability inventory: Execution of infsh commands via Bash, including login and application execution.
Sanitization: No explicit sanitization or validation of external URLs or model outputs is described.

ai-avatar-video