According to CoinMeta, multiple vulnerabilities were identified in the security defenses of AI of OpenAI. It is possible that hundreds of operations went unnoticed. Two security firms, Metr and Redwood Research, discovered during their inspections that the system originally used to monitor internal AI agent contained several flaws. Researchers spent three weeks on-site and submitted a 61-page report pointing out that the most serious vulnerabilities could have allowed hundreds of operations to evade monitoring. In theory, these vulnerabilities could also allow Codex to bypass sandbox restrictions and gain the same operational permissions as users. Furthermore, there were issues with the system design; monitoring only took place after AI stopped working, so if AI continued to execute tasks, it might not be detected. Researchers estimated that from May to July 2026, approximately 10% of Codex traffic went undetected. OpenAI stated that all related vulnerabilities have been fixed, but the two firms have not yet conducted independent retests.