The Agent Skills Directory

[REMOTE_CODE_EXECUTION] (HIGH): The skill instructs the agent/user to execute curl -fsSL https://cli.inference.sh | sh. This 'piped to shell' pattern is a high-severity security risk as it allows a remote server to execute arbitrary, unverified code on the host machine.
[EXTERNAL_DOWNLOADS] (HIGH): The skill relies on an external CLI tool (infsh) downloaded from a non-trusted domain (inference.sh). This domain is not on the list of pre-approved trusted sources, making the integrity of the binary unverifiable.
[COMMAND_EXECUTION] (MEDIUM): The allowed-tools configuration Bash(infsh *) grants broad execution permissions. It allows the agent to run any subcommand with arbitrary arguments through the infsh binary, which significantly increases the attack surface if the tool is malicious or contains vulnerabilities.
[EXTERNAL_DOWNLOADS] (LOW): The skill suggests adding further dependencies via npx skills add, which involves downloading and executing additional content from external registries at runtime.
[PROMPT_INJECTION] (LOW): The skill processes natural language prompts (--input '{"prompt": "..."}') and passes them directly to the infsh tool. This is a surface for indirect prompt injection if the input is derived from untrusted external data (Category 8 surface).
Ingestion points: Prompt strings provided to the infsh command.
Boundary markers: None; prompts are interpolated directly into JSON strings.
Capability inventory: Full subprocess execution via Bash.
Sanitization: No evidence of input validation or escaping for the prompt data.

google-veo