The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill handles external task inputs which are passed to multiple parallel subagents. Malicious content within these tasks could attempt to influence the subagents or the cross-judging process. \n- Ingestion points: User-provided tasks are defined in Phase A and distributed in Phase B (SKILL.md). \n- Boundary markers: The skill defines the task as the 'contract' but lacks explicit instructions to subagents to disregard embedded directives. \n- Capability inventory: The skill utilizes run_in_background for subagent spawning and performs file system operations in /tmp and git worktrees. \n- Sanitization: There is no evidence of task content validation or escaping before it is passed to subagents.

arena