The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The shell scripts convert.sh, format.sh, query.sh, and validate.sh all use heredocs (python3 << PYEOF) to execute Python code while directly interpolating shell variables like $FILE, $OUT, and $QUERY into the script body without escaping or validation. An attacker can provide a value containing a double quote (e.g., \"; import os; os.system('id'); #) to break out of the Python string literal and execute arbitrary Python code.
[COMMAND_EXECUTION]: The utility scripts generate and execute dynamic Python code at runtime based on unvalidated shell inputs, creating a high-risk execution path within the tool context.
[DATA_EXFILTRATION]: Due to the command injection vulnerability, an attacker can read sensitive files from the local file system (e.g., ~/.ssh/id_rsa, .env) and potentially exfiltrate their content if the environment allows network access.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it processes untrusted structured data from external files.
Ingestion points: The scripts convert.sh, format.sh, query.sh, and validate.sh read input data from files specified via the $FILE parameter.
Boundary markers: None. Output is printed directly to stdout without delimiters or 'ignore instructions' markers.
Capability inventory: The skill possesses file read/write capabilities and arbitrary code execution via the interpolation vulnerability.
Sanitization: None. The scripts perform no escaping or validation of file contents or query parameters before they are processed by the Python logic or interpolated into commands.

json-yaml