The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses shlex.split for safe command parsing and includes a validation script (scripts/validate_card.py) that proactively blocks shell command substitution (backticks, dollar-parenthesis) and dangerous redirection patterns in the model cards.
[DATA_EXFILTRATION]: The skill implements a strict privacy boundary. The validate_card.py script contains extensive regex patterns to detect and block the accidental inclusion of sensitive information such as IP addresses (IPv4 and IPv6), email addresses, absolute file paths, and scheduler job metadata.
[CREDENTIALS_UNSAFE]: The skill explicitly forbids and scans for hardcoded secrets, tokens (e.g., HF_TOKEN, ACCESS_TOKEN), and API keys within the verification cards through the SECRET_ASSIGNMENT_RE and SECRET_FLAG_RE detectors in the validation script.
[DYNAMIC_EXECUTION]: The static analysis warning regarding eval() in scripts/verify_hf_inference.py is a false positive. The code calls model.eval(), which is a standard PyTorch library method used to set a model to evaluation mode, not the dangerous Python eval() function.
[DYNAMIC_EXECUTION]: The yaml.load() call in scripts/validate_card.py is used with a custom _UniqueKeyLoader that inherits from yaml.SafeLoader. This ensures that only standard YAML scalars are processed, preventing unsafe object deserialization attacks.

create-model-verification-card