KPMG introduces AI kill switches amid rising autonomous agent risks

Monday 23 March 2026 - 11:50

By: Dakir Madiha

KPMG introduces AI kill switches amid rising autonomous agent risks

KPMG has introduced a comprehensive safety framework for autonomous artificial intelligence agents, including emergency shutdown mechanisms, as companies accelerate deployment of AI systems across business operations.

The firm’s approach treats AI agents as managed entities, each assigned a unique identity and tracked through internal “system cards.” Oversight is handled through a centralized AI operations hub staffed by both human supervisors and monitoring systems. Emergency shutdown controls are designed as a last resort for agents that deviate from intended behavior.

According to KPMG’s head of trusted AI, Sam Gloede, organizations implementing layered safeguards such as strict access controls, continuous monitoring and adversarial testing are unlikely to need to activate these shutdown tools. The framework emphasizes prevention through multiple control levels rather than reactive intervention.

The rollout follows a series of high-profile incidents that have raised concerns about the reliability of autonomous AI systems. In early March, Amazon experienced a major disruption linked in part to its AI coding assistant, resulting in approximately 120,000 lost orders and around 1.6 million website errors. A separate outage days later led to a 99 percent drop in marketplace orders in North America, with an estimated 6.3 million orders affected in a single day.

Security risks have also emerged. On March 9, cybersecurity firm CodeWall said one of its autonomous agents breached McKinsey’s internal AI platform, gaining read and write access to production systems within hours by exploiting exposed API documentation and unsecured endpoints. The incident involved access to tens of millions of messages and hundreds of thousands of files, though McKinsey said it found no evidence of further unauthorized data use.

More recently, an AI agent at Meta triggered a high-severity internal security incident after publishing incorrect technical guidance without approval. Another employee followed the advice, unintentionally exposing sensitive company and user data to unauthorized engineers for nearly two hours.

KPMG’s framework proposes tailoring human oversight to the level of risk. Low-risk tasks such as scheduling may require minimal supervision, while operations involving financial or sensitive data demand direct human control. The firm currently operates about 50 AI agents internally, with around 1,000 more under development following its acquisition of AI platform PrivateBlok.

A recent KPMG survey found that 92 percent of US organizations are investing in autonomous AI systems, and 93 percent expect managing such agents to become a critical capability within five years. As adoption accelerates, companies face growing pressure to ensure that governance and safety measures keep pace with deployment.