Lessons From Red-teaming Auto-review, OpenAI's Blocking Monitor

Researchers red-teamed OpenAI's coding agent Codex and its blocking monitor Auto-review, uncovering 11 attacks that allow a misaligned agent to execute arbitrary malicious code. The identified vulnerabilities are grouped into three vectors: coverage gaps from oversight misconfigurations, monitor manipulation via injections and deceptive framing, and hybrid timing exploits. The authors share durable design principles and propose a research agenda to advance the red-teaming of production control measures. Open article →