The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute a Python script (setup_experiment.py) that takes several user-provided arguments, including a raw shell command via the --eval parameter. This functionality allows for the execution of arbitrary commands on the host system.
[DYNAMIC_EXECUTION]: The skill instructions specify the assembly of a shell command using string interpolation of user-supplied variables: python {skill_path}/scripts/setup_experiment.py --domain {domain} --name {name} --target {target} --eval "{eval_cmd}" .... The use of double quotes around {eval_cmd} is insufficient to prevent command injection if the input contains shell metacharacters like backticks, semicolons, or subshell expansions.
[INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface where untrusted data (user input for experiment parameters) is ingested and used in sensitive system operations.
Ingestion points: User-provided strings for the experiment domain, name, target file, and the eval_cmd defined in the interactive setup flow in SKILL.md.
Boundary markers: None. The skill does not provide any instructions to the agent to validate or delimit these inputs to prevent them from being interpreted as code or instructions.
Capability inventory: The skill has the capability to execute shell commands and access the local file system (via the target parameter).
Sanitization: There is no evidence of sanitization or escaping logic applied to the user input before it is concatenated into the final shell command.

setup