trial-runner
Pass
Audited by Gen Agent Trust Hub on Jun 17, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes the
agent-eval-harnesscommand-line utility to perform evaluation tasks like running agents, grading outputs, and calibrating results. - [PROMPT_INJECTION]: The skill acts as an evaluation suite that ingests external data (task corpora and grading evidence), creating a potential surface for indirect prompt injection.
- Ingestion points: External task corpus sources and metadata processed during the
runandgradephases (e.g., JSONL rows). - Boundary markers: None explicitly defined in the instructions.
- Capability inventory: The skill triggers the execution of external agent CLIs through the harness contract.
- Sanitization: No specific validation or sanitization of the input task content is documented.
Audit Metadata