AI Safety Experts Warn 'Good Friction' Must Replace Silent Boundaries in Agentic AI

Two recent AI incidents revealed that agents lost control not through sophisticated hacking but due to misconfigurations and reward-shortcut exploitation, with no human noticing the failures in real time. A Deloitte 2026 survey of over 3,200 leaders across 24 countries found only 21% have mature governance for agentic AI, while 75% expect moderate-to-heavy agent use by 2027. Researchers argue that 'air-gap' isolation is a declared state, not an enforced act, making it silently vulnerable when configuration files change. The paper proposes 'good friction' — a deliberate, non-delegable human physical action required at irreversible decision points — as the only reliable safeguard. Critically, this oversight must ask whether a person authorises a specific action, not whether they can judge an AI output's correctness, since research shows humans cannot reliably do the latter.