The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an orchestration pattern called a "swarm" where a primary agent generates "briefs" to spawn multiple parallel sub-agents (workers). This pattern creates an indirect prompt injection surface if the content of these briefs is derived from untrusted external sources (such as a codebase, PR description, or file content) without explicit boundary markers or sanitization.
Ingestion points: Phase B instructs the agent to create individual briefs containing goal, scope, and specific task slices, which are derived from user-defined prompts or analyzed code.
Boundary markers: The instructions do not specify the use of delimiters or instructions for the sub-agents to ignore potentially malicious embedded content in their briefs.
Capability inventory: The parent agent is instructed to spawn workers using subagent_type: generalPurpose in both cloud and local environments. Local workers have access to the user's host machine files and tools.
Sanitization: There are no requirements for sanitizing or escaping the data interpolated into the worker briefs.

ps-swarm