The Agent Skills Directory

Sensitive Credential Access: The skill documentation includes examples that explicitly direct the agent to access sensitive authentication files at hardcoded paths, such as /home/foobar/authentication/auth.txt. Referencing or accessing secrets from fixed locations on the filesystem is a significant security consideration that may lead to the exposure of authentication tokens or private keys.
Dynamic Code Generation and Execution: The instructions encourage the agent to write and execute custom Python or Bash scripts to perform API-related tasks rather than using predefined tools. This pattern of dynamic code generation introduces a potential risk where generated logic might perform unintended or unauthorized operations if not carefully constrained.
External Content Integration: The skill provides examples of fetching and reading instruction files from remote GitHub repositories. Retrieving remote data introduces a surface where external content could influence the agent's behavior, especially if the source repository is not under the user's direct control.
Indirect Prompt Injection Surface: By guiding the agent to ingest untrusted data from external URLs and simultaneously granting the capability to execute generated code, the skill creates a potential surface for indirect prompt injection. This occurs when instructions embedded within the processed external data attempt to manipulate the agent's logic or downstream script execution.
Command Execution Capabilities: The skill is designed to interact directly with the Bash tool to run self-contained scripts. This broad capability to execute shell commands should be managed with least-privilege principles to prevent unauthorized system access or modifications.

creating-skills