The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/update_state.py executes arbitrary shell commands through the run_check function, which calls subprocess.run(cmd, shell=True). The commands are retrieved from the integrations.json configuration file, allowing for potentially dangerous system operations if the configuration is tampered with.\n- [PROMPT_INJECTION]: The skill architecture facilitates indirect prompt injection by instructing the agent to trust the system state file.\n
Ingestion points: The agent is instructed in SKILL.md to read brain/system_state.json before answering status-related questions.\n
Boundary markers: No delimiters or instructions to ignore embedded commands are present in the agent instructions or the generated JSON data.\n
Capability inventory: The skill includes file-writing capabilities and shell command execution via subprocess.run in scripts/update_state.py.\n
Sanitization: The script captures the raw output of system commands and writes it directly to the JSON file without any escaping, filtering, or validation, allowing malicious text from a monitored service to influence the agent.

system-state-injection