The Agent Skills Directory

[COMMAND_EXECUTION]: The skill dynamically constructs and executes a shell command in Step 4 by interpolating user-supplied strings (<prompt-spec>, <picked-models>) into a command-line template. This pattern is vulnerable to command injection if the input contains shell metacharacters (such as backticks, semicolons, or command substitutions) that are not properly sanitized before being passed to the shell.
[CREDENTIALS_UNSAFE]: The skill explicitly checks for and reads sensitive credential information from the local file system (~/.claude/.credentials.json) and environment variables (ANTHROPIC_API_KEY, GOOGLE_API_KEY). While this access is required for the skill's intended purpose of benchmarking against specific providers, it represents a high-privilege operation involving private keys and tokens.

benchmark-models