The Agent Skills Directory

[OBFUSCATION]: The file PersianTerms.txt contains hidden Unicode characters (Zero-Width Space, U+200B) throughout the Table of Contents and glossary sections.
Evidence: Hidden characters following page numbers (e.g., 5\u200b, 6\u200b) and potentially interspersed within terminology terms.
Risk: While these can be artifacts of professional typesetting or PDF extraction, zero-width characters are a common vector for steganographic prompt injection, allowing an attacker to hide instructions from human review and standard text scanners.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process user-provided text through multiple sub-agents, creating a significant attack surface for indirect prompt injection.
Ingestion points: The skill accepts a {source_text} variable from the user in SKILL.md which is then passed to six different sub-agents (diction-grammar, beauty-eloquence, modern-standards, bwc-style, back-translator, and oral-spiritual).
Boundary markers: The instructions interpolate the user input directly into sub-agent prompts using simple curly braces (e.g., SOURCE TEXT: {source_text}). There are no robust delimiters (such as XML tags or unique markers) or specific instructions for the sub-agents to ignore commands embedded within the text to be translated.
Capability inventory: The skill utilizes the TeamCreate and Task tools to spawn sub-agents; a successful injection could compromise the review process or influence the final synthesis step.
Sanitization: There is no evidence of sanitization or validation of the input text to prevent it from containing malicious instructions targeted at the agent orchestration workflow.

persian