The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions describe the use of the Bash tool to execute provided Python scripts (websearch.py, webfetch.py). These scripts take user-controlled inputs, such as search queries and URLs, as command-line arguments.
[INDIRECT_PROMPT_INJECTION]: The webfetch.py script fetches and extracts text content from arbitrary external websites to be processed by the agent. This creates a surface for indirect prompt injection if the fetched pages contain malicious instructions.
Ingestion points: The standard output of the scripts/webfetch.py utility, which provides extracted text to the agent context.
Boundary markers: The skill instructs the agent to record citations (URL, title, date) but does not mandate technical markers to delineate the untrusted content from the agent's instructions.
Capability inventory: The agent has access to the Bash tool and the provided search and fetch scripts to interact with the environment.
Sanitization: The webfetch.py script uses BeautifulSoup to strip potentially active or irrelevant HTML tags such as <script>, <style>, and <iframe> before returning content.
[CREDENTIALS_UNSAFE]: The websearch.py script facilitates the storage of a Serper.dev API key in a local file at ~/.serper_key. While the script attempts to secure the file using 0o600 permissions and includes instructions for the agent not to read the key file directly, the presence of secrets on the local filesystem is a security-sensitive practice.
[EXTERNAL_DOWNLOADS]: The provided scripts require external Python libraries including curl_cffi, beautifulsoup4, and httpx. The skill instructions direct the user or agent to install these dependencies from public registries.

web-search-techniques