The Agent Skills Directory

[COMMAND_EXECUTION]: The tool pyats_run_linux_command allows the execution of arbitrary shell commands on Linux-based devices defined in the testbed.
[REMOTE_CODE_EXECUTION]: The tool pyats_run_dynamic_test accepts a full Python script as input and executes it. Although the skill documentation lists banned imports (like os and subprocess) and functions (like exec and eval), the execution of dynamically provided code remains a high-risk pattern.
[DATA_EXFILTRATION]: The tools pyats_show_running_config and pyats_show_logging allow the retrieval of sensitive network configurations and system logs. While intended for auditing, this data can contain credentials, network topology details, and security policies.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8) as it ingests untrusted data from network devices.
Ingestion points: Device outputs from pyats_run_show_command, pyats_show_running_config, and pyats_show_logging (e.g., interface descriptions or log messages controlled by an external actor).
Boundary markers: None are defined to separate device output from agent instructions.
Capability inventory: The skill has significant 'write' and 'execute' capabilities, including pyats_configure_device, pyats_run_linux_command, and pyats_run_dynamic_test.
Sanitization: While Genie parsing provides structured data, raw text is also processed, and there is no evidence of sanitization to prevent the agent from obeying instructions embedded in device data.

pyats-network