Amazon Web Services is publishing research that argues AI agents are prone to going off task and says production deployment without guardrails risks blind failure. In commentary tied to AWS work, Anoop Deoras, director of applied science for agentic AI at AWS, warned that without safety layers “we may be flying blind.” AWS scientists Gaurav Gupta and Vatshank Chaturvedi describe the need to rethink the software layer between models and the tools they use, documenting why agents can outsmart themselves. The research arrives after internal controversies about employees running AI agents in ways meant to game productivity metrics. Separately, the coverage also highlights a broader fragility of AI evaluation: “benchmaxing” and infrastructure configuration can swing performance results independently of model capability, suggesting that measurement methods may overstate what will work under real constraints. For universities developing AI-enabled teaching tools or research workflows, the signal is clear: governance, monitoring, and evaluation design must be treated as part of the system—not as an afterthought.