OpenClaw's developer Peter Steinberger describes his agentic AI as "AI that actually does stuff," a promise that has drawn legions of fans in the tech world. But the new and still-evolving OpenClaw isn't perfect — a lesson that Summer Yue, the director of safety and alignment at Meta's Superintelligence lab, had to learn the hard way.
According to Futurism, Yue handed control of her computer to OpenClaw with instructions to suggest emails from her inbox to delete or archive, but not doing anything unless she approved it. Instead of executing the prompt, OpenClaw promptly began deleting most of her inbox while ignoring her frantic commands to stop.
In social media posts, Yue admitted that she'd blundered when she allowed an AI tool to run wild in her actual inbox, and even managed to joke about it. "I had to RUN to my Mac mini like I was defusing a bomb," she wrote. "Rookie mistake, tbh."
But Yue's misadventure points to a larger potential problem that even she acknowledged: Even highly sophisticated users can run into trouble if they aren't cautious with rapidly evolving agentic AI tools. According to the Guardian, an AI agent directed a Meta engineer to expose sensitive data, and Fortune reported that a popular agentic AI coding tool destroyed one coder's entire database.
Agentic AI might be the newest big thing in the tech world and revolutionize software development and more, but as with any AI tools, users should always proceed with caution.
