Domain Evaluation Harness
Warn
Audited by Gen Agent Trust Hub on Jul 7, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill guides the user to create and execute shell scripts (
harness.shandreport.sh) within the.hyperagents/domains/directory. These scripts are invoked with arguments that could potentially be manipulated if the local environment is compromised. - [COMMAND_EXECUTION]: The skill facilitates the execution of shell scripts whose paths are dynamically retrieved from the
.hyperagents/config.jsonconfiguration file, allowing for the execution of code from computed paths at runtime. - [PROMPT_INJECTION]: The skill defines an architecture that ingests untrusted task data and agent-generated predictions, creating an indirect prompt injection surface.
- Ingestion points: External task list JSON files (
TASK_LIST) and generatedpredictions.csvfiles are processed by the harness and reporter scripts. - Boundary markers: There are no instructions or examples provided for using boundary markers or delimiters to isolate untrusted task content from the agent's core instructions.
- Capability inventory: The system has the capability to execute arbitrary shell scripts (harness and reporter) and write evaluation reports to the file system.
- Sanitization: The skill does not describe or implement any sanitization or validation logic for the input task data or the collected agent predictions before they are processed by the shell scripts.
Audit Metadata