The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides commands for starting agents with the --dangerously-skip-permissions flag, which is an explicit attempt to bypass safety guardrails and permission controls.
[PROMPT_INJECTION]: The instructions direct the agent to bypass standard skill loading constraints (specifically the disable-model-invocation setting) by accessing another skill's instructions directly from the filesystem (~/.agents/skills/herdr-orchestration/SKILL.md) rather than using the platform's skill management tools.
[COMMAND_EXECUTION]: The skill includes shell commands that use the --approve-for-me flag, which reduces user oversight by instructing the worker agents to auto-approve actions. It also utilizes ps -o command= -p $PPID to inspect the host environment's execution parameters.
[INDIRECT_PROMPT_INJECTION]: The skill serves as an orchestration layer for external tasks while possessing high-privilege capabilities and lacking defined safety boundaries.
Ingestion points: Orchestrates work for multiple worker agents based on tasks provided in the session context.
Boundary markers: The instructions lack explicit delimiters or warnings to ignore embedded instructions in the tasks being orchestrated.
Capability inventory: Ability to execute shell commands and launch sub-agents with safety filters disabled (--dangerously-skip-permissions).
Sanitization: No evidence of input validation or sanitization is present for the data passed to the worker agents.

fable-herdr-orchestration