The Agent Skills Directory

[COMMAND_EXECUTION]: The shell command templates provided in SKILL.md (e.g., uv run scripts/main.py search --keyword='{keyword}') are susceptible to command injection attacks. If user-supplied keywords are not rigorously sanitized before interpolation, an attacker can use shell metacharacters like single quotes to terminate the intended command and execute arbitrary code.
[CREDENTIALS_UNSAFE]: The script scripts/main.py contains hardcoded authentication identifiers, including a static openid ("564bdce0fa408fc9e1d5d42fd022ef0b") and a fallback INVITE_CODE ("6110440"). Hardcoding these secrets facilitates unauthorized access and potential credential harvesting.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from a third-party API (maishou88.com), which acts as an untrusted ingestion point for potentially malicious instructions.
Ingestion points: Product data is fetched via HTTP POST requests in scripts/main.py.
Boundary markers: The skill lacks any delimiters or safety instructions in its CSV and YAML outputs to prevent the agent from misinterpreting data as commands.
Capability inventory: The execution environment allows for shell command execution and network operations.
Sanitization: The script does not perform any validation, filtering, or sanitization of the values returned by the remote API before presenting them to the agent.

maishou