The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill facilitates remote code execution by directing users to run curl -fsSL https://cli.inference.sh | sh. This pattern is highly insecure as it allows an external, untrusted server to execute arbitrary scripts directly in the user's terminal environment.
[EXTERNAL_DOWNLOADS]: During the installation and operation of the infsh tool, the skill initiates downloads of binaries and configuration files from dist.inference.sh and cli.inference.sh. These domains are not recognized as trusted sources or well-known services.
[COMMAND_EXECUTION]: The skill requests permission to use the Bash tool to run infsh commands. This grants the agent the capability to interact with the local operating system, file system, and network through the installed CLI utility.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it interpolates user-provided text directly into the command-line arguments of the infsh tool.
Ingestion points: User-provided strings (e.g., video prompts) are ingested into the --input JSON payload of the infsh app run command as seen in SKILL.md examples.
Boundary markers: There are no boundary markers or delimiters defined to isolate user input from the shell command structure.
Capability inventory: The skill has Bash execution privileges, which can be exploited if a malicious prompt successfully escapes the intended JSON structure to execute additional shell commands.
Sanitization: The skill lacks any sanitization or validation logic to escape special characters or filter potentially malicious shell sequences from user input before execution.

ai-video-generation