evaluating-llms-harness

Warn

Audited by Gen Agent Trust Hub on Jun 29, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes os.system within SKILL.md to automate evaluation tasks during training loops. This involves interpolating variables such as training steps directly into shell commands, which could be exploited if those variables are maliciously manipulated.
  • [REMOTE_CODE_EXECUTION]: The documentation explicitly instructs users to use the --allow_code_execution flag for benchmarks like HumanEval. This flag enables the runtime execution of code generated by the LLMs being tested, which is a high-risk operation when evaluating untrusted models.
  • [COMMAND_EXECUTION]: The references/custom-tasks.md guide demonstrates the implementation of custom Python utilities that utilize the subprocess module to execute and verify code outputs during the evaluation process.
  • [REMOTE_CODE_EXECUTION]: The skill provides instructions for loading and executing arbitrary Python logic via the !function YAML tag in custom task configurations, allowing for the execution of external script logic within the evaluation harness.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 29, 2026, 12:55 AM
Security Audit — agent-trust-hub — evaluating-llms-harness