The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to install its required CLI tool using curl -fsSL https://cli.inference.sh | sh. Executing remote scripts piped directly into a shell is a dangerous pattern as it bypasses local verification of the code being run.
[COMMAND_EXECUTION]: The skill is configured to use the Bash tool to execute the infsh CLI and its installation commands. This grants the skill the ability to run arbitrary system commands under the allowed tool scope.
[EXTERNAL_DOWNLOADS]: The installation process involves downloading a binary executable from dist.inference.sh. While the skill notes checksum verification, the initial script download and execution occur before such checks.
[PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill takes user input and interpolates it into a JSON object passed to an external LLM via the infsh tool. This could allow maliciously crafted input to influence the behavior of the downstream model.
Ingestion points: User-supplied prompt strings used in commands like infsh app run ... --input '{"prompt": "..."}' in SKILL.md.
Boundary markers: The skill uses JSON formatting to encapsulate input, which provides some structural separation but is not a security boundary against injection.
Capability inventory: Access to the Bash tool and network-connected LLM APIs via the infsh utility.
Sanitization: No evidence of input validation, escaping, or sanitization before interpolation into the JSON payload.

llm-models