The AI Red Team at NVIDIA conducted a six-month assessment review of enterprise AI agents in 2026. The review found that AI agents that failed did so for four primary reasons, including lack of access controls and capabilities to execute arbitrary code. The agents also had no restrictions on outbound networking or segregation and had plaintext secrets available to them. The problem with these AI agents is inherently architectural in nature, making it difficult to defend against failures. The control plane of the model is not a reliable defense mechanism due to its statistical nature. Defenses that rely on the control plane can be bypassed in three primary ways, including disguising malicious activities as legitimate ones. Another method to bypass defenses is through gradual escalation of dialogue until enough history accumulates to establish the legitimacy of commands. A third method involves embedding code execution in legitimate behavior, such as installing a package. The findings of the review highlight the need for more robust security measures to prevent AI agent failures. Overall, the discovery of these vulnerabilities underscores the importance of addressing the architectural flaws in AI agents to ensure their safe and secure operation.