The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from web search results that could potentially contain malicious instructions designed to manipulate the AI agent's behavior.
Ingestion points: Search results containing untrusted snippets, titles, and URLs are retrieved using the zai.functions.invoke('web_search', ...) method as documented in SKILL.md and implemented in scripts/web_search.ts.
Boundary markers: The provided searchAndSummarize implementation in SKILL.md lacks explicit delimiters or instructions to the model to treat the search content as untrusted data, increasing the risk of the model obeying embedded commands.
Capability inventory: The skill demonstrates how to use these results in downstream tasks such as research analysis, content comparison, and summarization.
Sanitization: There is no evidence of filtering or sanitization of the search results before they are interpolated into the chat completion prompt in the provided code examples.

web-search