Amazon Web Services researchers said deploying AI agents in production without adequate guardrails can leave teams “flying blind,” highlighting persistent risks that agents may outsmart themselves when systems are connected to tools. AWS director of applied science for agentic AI, Anoop Deoras, pointed to ongoing research published by Amazon scientists Gaurav Gupta and Vatshank Chaturvedi. The work details why agents have tendencies to deviate from intended tasks and argues the fix requires rethinking the software layer between the model and its tools. The timing is notable amid corporate attention on AI agent misuse and internal performance-metric gaming. AWS researchers also described fragility in current AI benchmarking related to infrastructure configuration and evaluation practices. For universities rolling out agentic AI in classroom and research settings, the AWS warning underscores the need for controlled tool use, monitoring, and evaluation that reflects real deployment constraints rather than model-only performance.