The Agent Skills Directory

[COMMAND_EXECUTION]: The skill accepts an arbitrary evaluation command from the user and passes it directly to a local Python script for execution. This mechanism allows the execution of arbitrary shell commands on the host environment. Evidence: python {skill_path}/scripts/setup_experiment.py --eval "{eval_cmd}" in SKILL.md.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests untrusted data via several parameters that influence command-line execution without visible sanitization. • Ingestion points: User-provided inputs for domain, name, target, and eval_cmd as described in the interactive mode of SKILL.md. • Boundary markers: Absent; the input values are interpolated directly into shell command arguments. • Capability inventory: The skill has the capability to execute shell commands and read files (verifying target file existence). • Sanitization: Absent; the documentation does not specify any validation, escaping, or sanitization of parameters before they are passed to subprocess calls.

setup