The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external inputs to define test steps, parameters, and server-side scripts. These inputs could contain malicious instructions designed to influence the agent's behavior during the test creation or execution process.
Ingestion points: Untrusted data enters the context through snow_create_atf_test, snow_create_atf_test_step, and the parameters field in sys_atf_test.
Boundary markers: There are no explicit boundary markers or instructions to the agent to ignore embedded commands within the test data or scripts.
Capability inventory: The skill can execute server-side scripts (snow_create_atf_test_step with server_script type), modify ServiceNow records (snow_create_atf_test), and execute tests (snow_execute_atf_test).
Sanitization: No sanitization or validation of the input strings or scripts is mentioned before they are passed to the tool functions.
[DYNAMIC_EXECUTION]: The skill facilitates the creation and execution of custom server-side JavaScript within the ServiceNow environment.
Evidence: The documentation highlights "Run Server Side Script" steps where users can provide custom ES5 code blocks that are executed by the ServiceNow platform as part of the test runner.
Risk: If the agent is provided with a malicious script as part of its task, it will facilitate the execution of that code within the ServiceNow instance.
[COMMAND_EXECUTION]: The skill allows the agent to execute actions within a ServiceNow instance through specialized tools.
Evidence: Tools like snow_create_atf_test and snow_execute_atf_test allow for record creation and state changes in the target system.
Risk: This capability, while central to the skill's purpose, requires careful oversight to ensure it is not used to perform unauthorized modifications to the ServiceNow environment.

atf-testing