The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches content from arbitrary web URLs through agentic_fetch.py and the --auto-fetch-top feature in agentic_search.py. This content is returned to the agent to provide research data. \n
Ingestion points: scripts/agentic_fetch.py and scripts/agentic_search.py (via the auto_fetch_top parameter). \n
Boundary markers: The skill lacks explicit boundary markers or "ignore instructions" delimiters when returning fetched markdown to the agent, potentially allowing embedded malicious instructions in web pages to influence agent behavior. \n
Capability inventory: The skill has network access via the httpx library and can write to the local file system (temporary directory) for session caching. \n
Sanitization: Content is not sanitized or filtered, as the tool aims for "100% fidelity" of the source markdown.\n- [DYNAMIC_EXECUTION]: Multiple scripts (including agentic_search.py, agentic_map.py, and agentic_fetch.py) use __import__(\"pathlib\") to programmatically determine their file location and adjust sys.path. This is used to facilitate imports of local shared modules like _http.py and _session.py when the scripts are run directly, and it does not involve executing untrusted input.\n- [DATA_EXFILTRATION]: The _session.py module persists search results, including queries and fetched web content, to the system's temporary directory under agentic-search/sessions/. While this is an intended feature for composing research steps, it means potentially sensitive research data is stored locally in plaintext JSON files.\n- [EXTERNAL_DOWNLOADS]: The skill makes network requests to api.x.ai, api.tavily.com, and api.firecrawl.dev to fulfill its research functions. These connections are to the official endpoints of the search and LLM providers used by the skill author.

agentic-search