The Agent Skills Directory

[COMMAND_EXECUTION]: The skill allows the execution of arbitrary shell commands through the --eval parameter, which is run during the baseline capture and evaluation phases. Additionally, user input from the --task and --eval parameters is interpolated directly into shell commands used to call internal scripts, which may lead to command injection if inputs are not properly sanitized.
[PROMPT_INJECTION]: The skill processes untrusted input from the --task parameter and potentially from evaluation outputs, passing it to agents without sanitization or boundary markers.
Ingestion points: Untrusted data enters via the --task and --eval arguments in SKILL.md and through evaluation results in Step 5.
Boundary markers: There are no explicit delimiters or instructions provided to agents to ignore potentially malicious content within the user-supplied task description.
Capability inventory: The skill is capable of executing shell commands and Python scripts, modifying the local filesystem (.agenthub/sessions/), and spawning parallel AI agents with tool-calling capabilities.
Sanitization: No evidence of input validation, escaping, or filtering of user-provided content is present in the skill definition or instructions.

run