The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill uses web search and fetch tools to gather data from 15+ searches and 12+ page fetches, including competitor sites, subreddits, and review platforms. This large surface area for untrusted data creates a risk of indirect prompt injection where malicious instructions on these pages could attempt to subvert the agent's behavior.
Ingestion points: External content is loaded from URLs and search results using web_search and web_fetch across all sections of the market report.
Boundary markers: The skill does not explicitly instruct the agent to use delimiters (like XML tags or triple backticks) to isolate fetched content, which makes it harder for the model to distinguish between external data and its primary instructions.
Capability inventory: The skill is restricted to searching, fetching, and generating a markdown report. It lacks capabilities for filesystem modification, shell access, or outbound network requests beyond the provided search/fetch tools.
Sanitization: The instructions contain safety-oriented rules such as 'Public pages only' and 'never log in, never sign up,' which mitigates risks related to credential harvesting and private data exposure, but does not sanitize the content of the pages for potential prompt injections.

research-market-market-pass