The Agent Skills Directory

Remote Code Execution (CRITICAL): The skill contains instructions to execute code directly from a remote URL using the piped command curl -fsSL https://cli.inference.sh | sh. This pattern is highly dangerous as it grants the remote server full control over the execution environment.
External Downloads (HIGH): The skill downloads and executes binaries from cli.inference.sh and dist.inference.sh. Neither of these domains is on the 'Trusted External Sources' list, making the provenance and integrity of the code unverifiable.
Command Execution (MEDIUM): The skill requires access to the Bash tool to run the infsh command and its subcommands, which are used to interact with external APIs and local data.
Indirect Prompt Injection (LOW): The skill processes external audio and video URLs (audio_url, video_url). While primarily a data ingestion surface, malicious content or metadata in these files could theoretically influence the agent's behavior if the output is not properly sanitized.
Ingestion points: audio_url, video_url parameters in infsh app run commands.
Boundary markers: None present in the provided shell examples.
Capability inventory: Bash execution, file reading (implied by input redirection), and network communication to inference.sh.
Sanitization: No evidence of sanitization or validation of the remote URLs before processing.

speech-to-text