The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted biological sequence data and patent records, which represents a potential attack surface for indirect prompt injection.
Ingestion points: The user-intake-protocol.md and sequence-search.md files define the ingestion of VH, VL, CDR, and construct sequences, alongside patent data fetched via MCPs.
Boundary markers: The workflow utilizes structured tech profiles and intake gates, although it lacks explicit delimiters to prevent the agent from obeying instructions that could be hidden within sequence listings or patent claim text.
Capability inventory: The skill uses advanced_patent_search, patent_briefing, and drug_asset MCPs, and executes a local Python script scripts/mab_fto_recall_estimator.py for search diagnostics.
Sanitization: The instructions explicitly advise against placing secrets or API keys in queries and logs. Furthermore, the output-contract.md mandates that generated HTML reports must be self-contained and free of external network calls, fonts, or CDNs.
[SAFE]: All external URLs and tools referenced belong to the author's official domain (patsnap.com) or represent well-known industry databases (INPADOC).
[SAFE]: The Python script scripts/mab_fto_recall_estimator.py is a utility for data normalization and search diagnostics; it uses only standard libraries and contains no network operations, subprocess calls, or dynamic code execution.

screen-mab-fto-risk-ls