The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection vulnerability surface identified. The skill ingests arbitrary user text describing systems or problems and incorporates it directly into agent tasks that have access to administrative capabilities.\n- Ingestion points: User input is accepted via command-line arguments in scripts/systems-thinking.mjs, examples/capacity-estimator.mjs, examples/decompose-system.mjs, examples/observability-rollout.mjs, and examples/tradeoff-check.mjs.\n- Boundary markers: The skill uses simple text prefixes (e.g., 'Problem:', 'System:', 'Requirements:') in the prompts. It lacks robust delimiters such as triple backticks or XML tags that would prevent the agent from confusing user input with system instructions.\n- Capability inventory: The agent is empowered to use high-impact platform tools including run_fleet_command, manage_playbook, manage_change, manage_scheduled_task, and agtPolicyEngine to configure and modify infrastructure.\n- Sanitization: No validation, escaping, or filtering is applied to user-supplied text before it is template-interpolated into the agent's task description.

rterm-systems-thinking