TikTok: Shou Zi Chew Warns About AI Agent Permissions

Users report incidents of AI agents publishing sensitive data or making unwanted purchases. In response to these issues, TikTok CEO Shou Zi Chew calls for a careful review of permissions and a framework for the information shared with these tools. He detailed his warnings during a conference in Singapore.

Agents Going Awry: From Audit Leaks to 20 Pounds of Chicken

Users report that their personal AI agents have sometimes taken unexpected initiatives. Shane Mac explains that his agent, tasked with sending him a monthly bank audit, accidentally published this document in his company's Slack chat. Other users mention agents producing incorrect information or misinterpreting the instructions received. In Houston, Austin Welch indicates that his Instinct agent mistakenly ordered 20 pounds of chicken for a meal with his partner.

Chew Calls for Control Over Permissions and Usage Limits

Shou Zi Chew, CEO of TikTok, believes that many users grant permissions to their AI agents without fully understanding what they are allowing. He recommends taking the time to understand how these agents work and carefully reviewing the permissions granted. Chew cites the example of agents having access to credit card information, which can make purchases exceeding the initial request. He emphasizes that some users can be overly enthusiastic in their use of these tools and insists on the need to know the limits of what is reasonable and acceptable.

Internal Framework at TikTok and the Rise of Personal Agents

The leader specifies that TikTok trains its teams not to include sensitive information, such as trade secrets or privileged legal data, in public exchanges with AI agents. He presented these recommendations at the DBS Asia Future Forward Conference in Singapore, focused on creating value through AI in Asia. Chew also mentioned using AI as a dietary coach and for answering medical questions. Meanwhile, personal AI agents are proliferating, with tools like Meta's Muse, SpaceXAI's Grok Bot, OpenAI's Dots, and Instinct's agents often highlighted for their ability to handle repetitive tasks such as bookings, audits, or purchases. However, reports of incidents are increasing as these tools become more widespread.