The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill requires the use of sudo for several low-level operations, including network traffic capture with tcpdump, package installation via apt, and advanced Layer 2 discovery with nmap.- [REMOTE_CODE_EXECUTION]: The instructions direct the agent to clone and execute code from unverified third-party GitHub repositories, specifically TacticalGator/modbuster and yadox666/MATRIX_Modbus_Attack_Tool.- [CREDENTIALS_UNSAFE]: The methodology explicitly includes searching for plaintext passwords and domain credentials within PLC project files (such as .s7p, .acd, and .rsp), historian databases, and Windows HMI/Engineering Workstation configurations.- [DYNAMIC_EXECUTION]: Multi-line Python scripts are frequently executed dynamically using shell heredocs (`python3
<<'EOF'), a pattern used for various protocol interrogation tasks throughout the skill.- **[COMMAND_EXECUTION]:** The skill includes commands that directly interact with industrial hardware in ways that could be physically disruptive, such as forcing digital output states (write_coil) or halting PLC execution (plc_stop).- **[INDIRECT_PROMPT_INJECTION]:** The skill is susceptible to indirect prompt injection due to its processing of external industrial data. **Ingestion points**: Network traffic logs (Zeek/tcpdump) and protocol register/tag data (Modbus/S7/CIP). **Boundary markers**: Absent; there are no instructions provided to the agent to disregard potential commands embedded in analyzed traffic or data. **Capability inventory**: Broad network access, file system access, and command execution capabilities are available across the included scripts. **Sanitization**: Absent; variables such as and<hex_payload>` are interpolated directly into shell and Python commands without validation.

ics-technique