The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/serial_monitor.py utilizes subprocess.run() to execute the openocd binary for device reset functionality. Additionally, the _read_openocd_path() function retrieves the path to the executable from a local workspace file named .em_skill.json. If the agent runs this skill within an untrusted or maliciously crafted project directory, it could result in the execution of an unintended or malicious binary specified in the configuration file.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from external serial interfaces, presenting an attack surface for indirect prompt injection.
Ingestion points: In scripts/serial_monitor.py, external data is ingested from serial connections via ser.read() and printed line-by-line to the standard output.
Boundary markers: Absent. The captured serial text enters the agent's context without any encapsulation or structural delimiters.
Capability inventory: The skill contains system command execution capabilities via subprocess.run() inside scripts/serial_monitor.py.
Sanitization: The script performs text cleanup via strip_ansi() to remove terminal escape sequences, but it does not sanitize or validate natural language instruction content, which could allow a connected serial device to output malicious instructions targeting the agent.

serial-monitor