The Agent Skills Directory

[PROMPT_INJECTION]: The script scripts/web-search.sh takes the user-provided search query and interpolates it directly into a prompt string for the gemini command. This creates a direct injection vector where a user can provide instructions that override the intended behavior of the tool.
[COMMAND_EXECUTION]: The gemini command is executed with the --yolo flag. In many AI CLI tools, this flag permits the model to execute suggested shell commands or perform system actions without manual confirmation. When combined with the prompt injection vulnerability mentioned above, an attacker could trick the tool into executing harmful system commands (e.g., file deletion or credential theft).
[INDIRECT_PROMPT_INJECTION]: This skill possesses a significant attack surface for indirect injection.
Ingestion points: The $SEARCH_QUERY variable in scripts/web-search.sh accepts untrusted data.
Boundary markers: There are no delimiters or instructions used to separate the user query from the system's instructions.
Capability inventory: The skill executes the gemini CLI with automation permissions (--yolo).
Sanitization: No sanitization or validation is performed on the search query before it is passed to the command line.

searching-web