The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted external data from customer reviews and CRM notes, which could contain malicious instructions meant to influence the agent's behavior.
Ingestion points: The skill reads data from external sources including payment gateways (Yookassa, Tinkoff), CRM systems (Bitrix24, amoCRM), and uploaded review files (CSV/Excel) from sites like Yandex Maps, 2GIS, and Avito.
Boundary markers: The instructions in SKILL.md do not include explicit delimiters or specific warnings to the agent to disregard instructions embedded within the processed feedback data.
Capability inventory: The skill utilizes WebFetch for network access, Read for local file access, and Bash for general command execution, creating a significant attack surface if an injection occurs.
Sanitization: There are no documented sanitization or filtering steps to strip potentially malicious content from the feedback before it is analyzed by the agent.

customer-pulse-check