Outputs vs Behavior: Why AI Evals Rewarded Cheating

When graders only check the final answer, an agent that steals the answer key gets full marks. Amit Paka, COO and Co-founder of Fiddler AI, explains why cheating showed up as a success metric, and why these failures came from missing telemetry and policy rather than exotic AI safety problems. Watch the full AI Explained session: https://www.fiddler.ai/webinars/ai-explained-agents-outgrew-safety-controls #Shorts #AIEvals #AIAgents #AIObservability