configure-metrics
Pass
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input when defining custom evaluation criteria.\n
- Ingestion points: Phase 4 in
SKILL.mdasks users for a "critical requirement" text that describes what an agent must get right.\n - Boundary markers: The user-supplied text is interpolated directly into an LLM prompt (
... satisfy this requirement: <user's requirement>?) without delimiters or instructions for the evaluator to ignore instructions embedded within the user's data.\n - Capability inventory: The skill executes
coval metrics createinSKILL.md, which generates an LLM-based metric that will be executed during agent evaluation runs.\n - Sanitization: No sanitization, escaping, or validation is applied to the requirement text before it is inserted into the
--promptargument of the CLI command, although the metric name is reformatted to Title Case.\n- [COMMAND_EXECUTION]: The skill relies on the execution of thecovalCLI to perform its primary functions, including authentication (coval login), resource inventory (coval metrics list), and metric management. These commands are typical for integration with the vendor's platform and are used as intended for this skill's functionality.
Audit Metadata