The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions include examples of using the Bash tool to run project-specific tests (npm test) and file content checks (grep). These operations are intended to verify the success of a development task and are performed locally within the user's project environment.
[DATA_EXPOSURE]: The framework defines a local storage structure for evaluation results and logs in .claude/evals/. There is no evidence of instructions to read sensitive system files (e.g., SSH keys, environment variables) or to transmit data to external servers.
[PROMPT_INJECTION]: The skill uses instructional headers and markers (e.g., [CAPABILITY EVAL]) to structure evaluation criteria. It does not attempt to override the underlying agent's safety guidelines or system instructions.

eval-harness