The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from an external task management system (Iskron), which exposes the agent to potential indirect prompt injection attacks.
Ingestion points: The skill retrieves content from external 'cases' and 'graphs' using specialized platform tools like iskron_case(action="read") and iskron_orient(lens="board") as detailed in the 'Сводка: найти, прочесть, свести' section of SKILL.md.
Boundary markers: The skill explicitly instructs the agent to use strict formatting for external content, using 'цитатой с провенансом' (quote with provenance) and enclosing external text in quotes (e.g., №12 · rN · [340] · кто сказал · когда — «слова записи как есть»), which serves as a delimiter to separate task data from core instructions.
Capability inventory: The skill possesses significant write capabilities, including sending messages (iskron_case(action="say")), inviting roles to cases (iskron_case(action="invite")), and updating task lines (iskron_case(action="line")), which could be exploited if malicious instructions inside a task were followed.
Sanitization: No automated filtering or sanitization of input data is mentioned; the skill relies on the agent's adherence to the specified formatting and provenance rules.

assistant