The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's core functionality includes submitting and managing shell jobs via the gbrain CLI, allowing for arbitrary command execution. The documentation explicitly warns that enabling this feature (via GBRAIN_ALLOW_SHELL_JOBS=1) authorizes remote code execution and must be treated as a privileged authorization.\n- [INDIRECT_PROMPT_INJECTION]: The skill supports spawning subagents and steering their behavior mid-flight through inbox messages containing directives. This creates an attack surface where external input could influence the agent's reasoning loop or cause it to deviate from its intended tasks. Evidence Chain: Ingestion points include the gbrain agent run prompt and send_job_message payload in SKILL.md; boundary markers are present in the form of structured JSON directives; capabilities include shell command execution and subagent tool usage; explicit sanitization steps are not detailed in the instructions.\n- [PERSISTENCE]: The deadman check pattern for durable execution involves creating one-shot or recurring scheduled tasks using host tools like cron, at, or crontab. While used for legitimate monitoring and error recovery of long-running jobs, these mechanisms represent a form of persistent task scheduling on the host machine.\n- [REMOTE_CODE_EXECUTION]: The subagent orchestration capability enables spawning agents that can utilize various research and synthesis tools. This involves complex reasoning loops and network operations that could be exploited if subagents are given malicious instructions.

minion-orchestrator