The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_selection_eval.py invokes the claude CLI tool using subprocess.run with the --permission-mode bypassPermissions flag. This flag suppresses standard security confirmations that usually require a human to approve tool execution (e.g., file writes, command runs). This automation bypasses a critical security control and could lead to unauthorized actions if the script is run with malicious inputs.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by processing prompts from evals/evals.json and feeding them directly into an elevated execution environment.
Ingestion points: The prompt and context fields in evals/evals.json.
Boundary markers: None. The prompts are interpolated directly into the CLI command without delimiters or instructions to ignore embedded commands.
Capability inventory: The target environment (claude with bypassed permissions) has full access to available agent tools, including file system and shell access.
Sanitization: No validation, escaping, or sanitization of the evaluation prompts is performed before they are passed to the CLI.

skill-selection-evals