The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external aggregators and government portals, which creates a surface for indirect prompt injection if the source data contains malicious instructions.
Ingestion points: Data is retrieved from bo.nalog.gov.ru, egrul.nalog.ru, and pb.nalog.ru via scripts/fetch_counterparty.py, and from checko.ru via browser actions.
Boundary markers: The SKILL.md file includes a 'Safety-floor' section that instructs the agent to treat external content as data and to ignore any embedded commands.
Capability inventory: The skill uses network requests to fetch data; it does not have capabilities for arbitrary command execution or writing to the file system.
Sanitization: INN inputs are validated using checksums and length checks. Sanitization of external content is handled via natural language instructions in the skill's prompt context.

counterparty-guard