The Agent Skills Directory

[METADATA_POISONING]: The skill metadata lists 'arize' as the author, which conflicts with the provided author context identifying 'ethan-rio' as the author. This discrepancy is a form of deceptive metadata.
[INDIRECT_PROMPT_INJECTION]: The skill provides commands such as ax datasets create and ax datasets append that ingest untrusted data from various file formats and strings. The evidence chain includes: 1) Ingestion points for external data via the --file and --json flags; 2) A lack of explicit boundary markers or instructions to ignore instructions within that data; 3) Capabilities to execute CLI commands and interact with the filesystem; 4) No mentioned sanitization or validation of the data contents. This surface allows for potential multi-step attacks if malicious data influences the agent's behavior.
[PERSISTENCE]: In references/ax-profiles.md, the skill instructs users on how to persist environment variables by modifying shell profiles like ~/.zshrc or ~/.bashrc. While this is intended for legitimate configuration of the ARIZE_SPACE variable for the CLI tool, it represents a mechanism for maintaining persistent configuration across sessions.

arize-dataset