The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local JavaScript runner (evals/runner.js) using node. It utilizes child_process.spawnSync with an argument array (argv) rather than a shell string, which prevents command injection by ensuring arguments are not interpreted as shell commands.
[INDIRECT_PROMPT_INJECTION]: The skill accepts a skill name as an argument to determine which evaluation to run. It mitigates potential injection risks by enforcing a strict regular expression validation (/^[a-z][a-z0-9-]{0,63}$/), effectively blocking shell metacharacters, spaces, and directory traversal characters.
[DATA_EXPOSURE]: Evaluation results are persisted to the .bkit/runtime/ directory. The file paths are constructed using the validated skill name and an ISO timestamp, ensuring that the tool cannot be used to write to arbitrary locations on the file system.

bkit-evals