The Agent Skills Directory

Remote Code Execution (CRITICAL): The skill explicitly commands the execution of a remote script from https://cli.inference.sh by piping it directly to the shell (| sh). This pattern bypasses verification and allows an external source to execute arbitrary code on the user's system.\n- External Downloads (MEDIUM): The skill references and installs additional capabilities via npx skills add inference-sh/skills, which points to an untrusted third-party organization not included in the safe list.\n- Command Execution (MEDIUM): The skill utilizes Bash(infsh *), granting the agent broad authority to execute any subcommand of the potentially malicious infsh CLI tool downloaded in the previous step.\n- Indirect Prompt Injection (LOW): The skill provides an interface for processing external text through the infsh tool without sanitization.\n
Ingestion points: User input is interpolated into the prompt field of the infsh app run command within SKILL.md.\n
Boundary markers: None present to isolate user-provided text from command instructions.\n
Capability inventory: Uses bash to interact with external inference APIs and manage local file output.\n
Sanitization: No input validation or escaping mechanisms are implemented for the processed text.

dialogue-audio