The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted search result data (smart snippets and snippets) retrieved from the Yandex Search API. This content could potentially contain malicious instructions intended to subvert the agent's behavior.
Ingestion points: Data enters the agent's context through scripts/web_search_sync.sh and scripts/web_search_async.sh via the Yandex Search API endpoints.
Boundary markers: The skill organizes results into a markdown pack (render_snippet_pack in scripts/common.sh) using markdown headers (##) and separators (---), but it lacks explicit instructional warnings to the agent to disregard any commands embedded within the fetched text.
Capability inventory: The skill allows the agent to execute local shell scripts, perform network requests via curl, and write/read files within the project's cache/ directory.
Sanitization: While the skill correctly parses the API's JSON and XML responses using Python's standard libraries to prevent structural injection, it does not sanitize the semantic content of the search results for prompt injection patterns.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to official Yandex Cloud API domains (*.api.cloud.yandex.net) for authentication and search operations. These are well-known services being used according to their intended functionality.
[COMMAND_EXECUTION]: Several shell scripts execute Python 3 one-liners to perform data transformation and hashing. The skill mitigates command injection risks by passing all external data (such as search queries and file paths) to Python via environment variables (os.environ) rather than interpolating them directly into script strings.

yandex-search-api