The Agent Skills Directory

[DATA_EXFILTRATION]: The documentation in references/detailed-guide.md includes hardcoded absolute file paths (e.g., C:\Users\renat\skills\agent-orchestrator\scripts\scan_registry.py) in its implementation examples. This discloses the author's local system username and directory structure, which is an unnecessary exposure of internal developer environment details.\n- [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process untrusted user-supplied data (criminal case descriptions) for legal analysis. It does not provide the AI agent with specific instructions to use delimiters (such as XML tags or triple quotes) or to ignore potential commands embedded within that data. This creates a surface for indirect prompt injection, particularly given the agent's access to powerful tools like claude-code and cursor.\n
Ingestion points: User-supplied case descriptions processed during the 10-step workflow in references/detailed-guide.md.\n
Boundary markers: Absent. The skill provides no delimiters for external content.\n
Capability inventory: The skill is configured to use tools such as claude-code, cursor, gemini-cli, and codex-cli, which have capabilities for file system modification and command execution.\n
Sanitization: No instructions for sanitizing or validating user input are provided.\n- [METADATA_POISONING]: The skill includes self-authoritative safety claims such as 'Auditoria: Validada por skill-sentinel' and 'risk: safe' in the frontmatter and documentation. These claims are designed to influence or bypass independent security reviews by asserting safety and validation that cannot be externally verified.

advogado-criminal