The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests external data from the user, such as descriptions of foreign gifts, trust distributions, and donor relationships, which are then interpolated into a structured tax draft. This creates a surface for indirect prompt injection if the user-supplied data contains malicious instructions intended to mislead the agent during document generation.
Ingestion points: User-supplied gift and trust data described in the 'Prerequisites' and 'Workflow' sections of SKILL.md and detailed in references/line-by-line.md.
Boundary markers: The skill does not define explicit delimiters (e.g., XML tags or special markers) to isolate user-supplied strings from the agent's instructions.
Capability inventory: The skill is documentation-based and does not include executable scripts, network capabilities to untrusted domains, or file system writing operations beyond generating the text draft.
Sanitization: No explicit instructions are provided to sanitize, escape, or validate the content of user-provided strings before they are used to populate the draft tables.
[SAFE]: The skill follows security best practices for handling sensitive Personal Identifiable Information (PII). It explicitly instructs the agent to never store SSNs, ITINs, or EINs in logs or vector stores and to discard them immediately after use in the filing process.

form-3520