The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The instructions in SKILL.md (Phase A and Phase C) direct the agent to programmatically set the modeId of subagents to full-access or bypassPermissions. While the skill includes natural language constraints for these agents to remain "read-only," the use of these configuration flags bypasses the platform's default restricted permission sets for spawned agents.
[INDIRECT_PROMPT_INJECTION]: The skill uses a fan-out and synthesis pattern where user-provided tasks are processed by parallel subagents, and their outputs are subsequently analyzed by a "Judge" and the parent agent. This creates a vulnerability surface where a candidate's output could potentially influence the judge or synthesis process through embedded instructions.
Ingestion points: User-provided task prompt (Phase A).
Boundary markers: Absent; the instructions do not specify using delimiters or "ignore instructions" tags when passing candidate rationales to the Judge or during the synthesis Phase E.
Capability inventory: create_agent, list_profiles, list_models.
Sanitization: Absent; the skill directly processes text proposals and rationales from subagents.

arena