The Agent Skills Directory

[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill requires the installation of external tools via npx skills add, specifically targeting belt-sh/cli and other inference-sh resources. These are necessary components for integrating with the inference.sh platform.
[COMMAND_EXECUTION]: The skill uses a restricted bash environment through allowed-tools: Bash(belt *), which limits the agent to running commands related to the belt CLI tool, following the principle of least privilege.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the retrieval of external data from the web (via Tavily and Exa apps) and passes it to LLMs, which is a known attack surface for indirect prompt injection.
Ingestion points: Web content retrieved through apps like tavily/extract and exa/extract is directly used in LLM prompts for summarization and research tasks (SKILL.md).
Boundary markers: The provided examples do not include delimiters or specific instructions to help the model distinguish between system instructions and potentially malicious content within the extracted data.
Capability inventory: The agent uses the belt CLI to perform searches and invoke AI models, restricted to the belt command space.
Sanitization: There is no evidence of content sanitization or validation of the external web data before it is processed by the AI models.

web-search