The Agent Skills Directory

[REMOTE_CODE_EXECUTION] (CRITICAL): The skill instructs the user to execute a remote script via a pipe to a shell, which is an extremely dangerous pattern.
Evidence: curl -fsSL https://cli.inference.sh | sh found in SKILL.md.
Risk: This executes arbitrary code from a non-trusted domain (inference.sh) with the user's current privileges, allowing for potential malware installation or data theft.
[EXTERNAL_DOWNLOADS] (HIGH): The skill promotes the installation of additional unverified components from untrusted sources.
Evidence: Multiple instances of npx skills add inference-sh/skills@... in the Related Skills section.
Risk: Encourages expanding the attack surface by downloading and running more third-party code without integrity verification.
[COMMAND_EXECUTION] (MEDIUM): The skill relies on executing a third-party CLI tool with complex, user-generated inputs.
Evidence: Extensive use of the infsh tool to process JSON payloads containing prompts and code.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill demonstrates patterns that are vulnerable to indirect prompt injection when processing external data.
Ingestion points: The --input JSON payload in infsh app run commands (SKILL.md).
Boundary markers: Absent; raw strings are used within JSON structures.
Capability inventory: Shell access (via the infsh utility).
Sanitization: Absent; the examples show direct interpolation of potentially untrusted code snippets into LLM prompts.

prompt-engineering