The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted external data, such as customer-provided documents (PDFs/scans) and news search results, which creates an attack surface for indirect prompt injection. Malicious actors could embed instructions in these documents to bias the agent's risk assessment or hide negative findings.\n
Ingestion points: The agent uses a doc-reader tool to extract data from external files (PDF/scans) and ingests content from external search engine results (Baidu, Sogou) for news screening.\n
Boundary markers: The skill instructions explicitly include a 'Document Untrustworthiness' principle and require cross-verification from multiple sources, which provides a conceptual safeguard but does not implement technical delimiters for the processed content.\n
Capability inventory: The skill evaluates KYC rules, performs entity extraction, and generates risk reports. It does not appear to have access to sensitive system tools, shell execution, or arbitrary file writing that would enable high-severity exploitation.\n
Sanitization: The security constraints emphasize evidence-based reporting and manual review, but there is no mention of technical sanitization, escaping, or filtering of the raw text ingested from untrusted sources.

kyc-screener