The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill includes defensive instructions to prevent remote code execution (RCE) by advising against object deserializers that execute constructors or type callbacks. It recommends the use of data-only formats like JSON, MessagePack, or FlatBuffers validated against a schema.
[DATA_EXFILTRATION]: It proactively identifies and mitigates path traversal risks by instructing the user to sanitize profile and slot names against a strict allowlist character set before path construction.
[INDIRECT_PROMPT_INJECTION]: The skill acknowledges that save files are untrusted input and provides a detailed evidence chain for security:
Ingestion points: External save files from user data directories (user://, Application.persistentDataPath).
Boundary markers: Uses cryptographic integrity verification (SHA-256/HMAC-SHA256) and magic bytes to validate boundaries.
Capability inventory: Uses file system tools (write_file, read_file, create_file) specifically for persistent storage.
Sanitization: Instructs on validating deserialized fields against schema types and ranges before hydration.
[SAFE]: The skill implements industry-standard safety protocols for file I/O, including atomic write-and-swap operations (using .tmp and .bak files) and monotonic schema versioning to prevent data loss or corruption.

implement-save-persistence