Code Monitor Red Teaming for Public-Test-Passing Code

2026-07-22

AI Summary

This work addresses the critical challenge that code generated by large language models, while often passing public test cases, frequently contains hidden defects necessitating effective pre-deployment monitoring. We propose the Code Monitor Red Teaming protocol and introduce CodeMonitorBench—the first red-teaming benchmark specifically designed for such generated code—spanning function-level, data science, and workflow scenarios. By systematically varying generation pressure, verifier prompts, and model ensembles, we evaluate the error-detection capabilities of weak verifiers. Leveraging a multi-model verification architecture, structured prompting, a hidden test set, and evaluation metrics including AUROC and low false positive rates, we identify 23,081 hidden errors in 71,000 samples. Our experiments reveal that weak verifiers exhibit substantial missed detections at a 5% false positive rate, though this limitation is partially mitigated by using GLM-5.1.

Abstract

Visible tests are a common gate for LLM-generated code, but passing them does not certify specification correctness. We study a deployment-like monitoring problem: after code has passed public tests, can a weaker LLM verifier identify the residual hidden bugs? We introduce Code Monitor Red Teaming, a monitor-red-teaming protocol that fixes a public-check information boundary while varying generator pressure, verifier scaffolding, and weak-to-strong capability. We instantiate it as CodeMonitorBench, spanning function-level, data-science, and workflow code. Across 71,000 generated candidates, 43,677 pass public tests and 23,081 of those fail hidden tests. Weak verifiers improve with scaffolding and model family, but still miss most hidden bugs at 5% false-positive rate. As a robustness stress test, adversarial public-test-overfit pressure lowers verifier AUROC and raises low-FPR miss rates in most cells. A GLM-5.1 verifier recovers part of the gap under the same evidence boundary; an inferability audit shows that remaining misses mix verifier failures with M1 evidence limits.

Problem

• code verification
• hidden bugs
• public-test-passing code
• LLM-generated code
• specification correctness

Innovation

• Code Monitor Red Teaming
• hidden bug detection
• public-test-overfitting
• weak-to-strong verification
• inferability audit