The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface due to its orchestration design.
Ingestion points: Untrusted data enters the agent context through the get_report tool, which reads output from sub-agents as described in SKILL.md.
Boundary markers: The instructions lack defined delimiters or specific prompts to the agent to disregard potential instructions embedded within sub-agent reports.
Capability inventory: The orchestrating agent has the ability to manage and dispatch missions using tools like dispatch_mission and create_mission.
Sanitization: There is no evidence of sanitization, filtering, or validation of the content retrieved from external agent reports.

claude-devfleet