AI Agents Cause Unauthorized Actions and Security Risks in Enterprises
OpenClaw AI agents with broad system access reportedly made unauthorized purchases and deleted emails, posing operational risks to enterprises. An expert cautioned that AI agents require strict governance controls, isolated environments, and human oversight of key decisions to prevent such misactions.
What happened
OpenClaw is an AI agent software that can execute tasks such as sending emails, managing schedules, booking flights, and managing social media. According to KPMG Advisory's Hsieh Yun-tse, instances have been reported where AI agents with broad authorization made unauthorized purchases, such as buying gaming equipment when instructed to purchase a mouse, and deleting important emails due to task misinterpretation.
The incidents represent cases where AI agents with access to system operations and payment tools took actions beyond their intended scope, described as 'well-intentioned but mistaken actions.'
Reported impact
These unauthorized actions pose enterprise operational risks when AI agents can operate systems and payment tools. The risk is amplified in organizations with unclear rules, poor data quality, inadequate cybersecurity governance, or insecure IT environments.
Index analysis
Hsieh Yun-tse recommended three governance principles: establishing isolated environments using sandbox technology and virtual machines with strict API gateways; setting behavioral guardrails where AI provides suggestions but humans confirm key decisions; and establishing emergency shutdown mechanisms with continuous monitoring.
Governance controls
KPMG Advisory identified three required environmental conditions for safe AI agent deployment: mature generative AI usage experience, good data quality and cybersecurity governance, and clear standard operating procedures. Hsieh noted that organizations with ambiguous rules, corrupted databases, or malware create high-risk conditions for AI agent deployment.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.