The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content, such as 'benchmark samples', 'third-party content', and 'screenshots', to perform structural analysis or check for 'AI-likeness'. This creates an ingestion surface where malicious instructions could be embedded in the analyzed text.
Ingestion points: Found in SKILL.md, including user-provided benchmark copies, screenshots, and third-party text samples.
Boundary markers: The skill explicitly references an 'external material safety boundary' protocol (06_external-material-safety) which is triggered when processing external content.
Capability inventory: The skill possesses significant capabilities, including the ability to 'pack Eva data to desktop' and 'backup all memory cards', implying file system write access.
Sanitization: The skill relies on external safety protocols for sanitization rather than internal logic within the provided files.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill includes functionality to 'Export Eva data', 'Backup all Eva memory cards', and 'Pack Eva data to desktop'. While these are intended features for data management, they represent a high-privilege data movement surface.
Evidence: Triggers such as 'Export Eva data' and 'Backup all Eva memory cards' are documented in SKILL.md.
Mitigation: The instructions mandate that any export or backup action must be explicitly confirmed by the user and prohibits automatic authorization during inventory tasks.

eva-think