The Agent Skills Directory

[DYNAMIC_EXECUTION]: The tool pyats_run_dynamic_test enables the execution of dynamic Python code provided as a string at runtime.
Evidence: The skill body documents the pyats_run_dynamic_test tool which takes test_script_content (Python AEtest scripts) and executes them.
Risk: Although the instructions list a blacklist of banned imports (e.g., os, subprocess) and functions (e.g., eval, exec), blacklists are a fragile security control that can often be bypassed to achieve arbitrary code execution.
[COMMAND_EXECUTION]: The pyats_run_linux_command tool allows the agent to execute shell commands on remote devices.
Evidence: The tool is explicitly designed to execute shell commands on Linux-based devices within the network testbed.
Risk: This provides a direct path for the agent to execute arbitrary commands on infrastructure if proper authorization and validation are not enforced by the underlying MCP script.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from network devices which could contain malicious instructions designed to influence agent behavior.
Ingestion points: Device output is collected via pyats_run_show_command, pyats_show_logging, and pyats_show_running_config (SKILL.md).
Boundary markers: The skill instructions do not specify any boundary markers or delimiters (e.g., XML tags) to separate device output from agent instructions.
Capability inventory: The skill possesses write capabilities (pyats_configure_device), remote execution (pyats_run_linux_command), and dynamic code evaluation (pyats_run_dynamic_test).
Sanitization: No sanitization or filtering of device-provided text (such as interface descriptions or log messages) is mentioned before the data is returned to the agent context.
[DATA_EXFILTRATION]: The skill provides tools to retrieve sensitive device information that could be exfiltrated if the agent is compromised.
Evidence: The pyats_show_running_config tool retrieves the complete configuration of network devices, which typically contains sensitive data such as password hashes, SNMP strings, and internal network topology details.

pyats-network