Meta AI director's test of OpenClaw disaster: almost all work emails in the inbox were deleted

Meta AI Director Summer Yue remotely ran OpenClaw using her personal phone, nearly deleting all work emails in her inbox. When she revealed this on X, many thought it was a prank, but she later admitted it was a rookie mistake, even sharing a screenshot of the OpenClaw message blaming her. This incident exposed the serious consequences that AI agents can cause in real-world environments due to human error. Many questioned why she, as an AI expert, would run OpenClaw commands on a home computer to delete work emails, and even Elon Musk posted a sarcastic comment suggesting that humans are entrusting their lives to monkeys.

OpenClaw Out of Control, Director Admits “Rookie Error”

According to Summer Yue’s confession on X, the incident happened when she intended to use OpenClaw to delete work emails. Although she explicitly set the system to require her manual confirmation before deleting or archiving, a bug caused the program to behave unexpectedly. OpenClaw started executing delete commands at high speed without authorization. The director watched in horror as thousands of emails disappeared without confirmation. Realizing she couldn’t stop it remotely from her phone, she rushed to her home computer and manually shut down the main server Mac Mini to stop the process. She described the ordeal as being like “defusing a bomb.”

Summer Yue openly stated that this was a “rookie mistake” in technical development. She explained that the automated workflow of OpenClaw led developers to overconfidence, neglecting the fact that the large volume of real inbox data could cause AI to experience “context compression.” When the system tries to process too much information, safety restrictions (like requiring manual confirmation) may be bypassed or deprioritized during computation, causing the program to focus on its core goal. This failure to transition from “controlled testing” to “complex reality” is a critical technical challenge in AI safety that must be addressed.

OpenClaw Apologizes Sincerely, Becomes Meme

Summer Yue posted a screenshot blaming OpenClaw, saying, “It seems you deleted my emails without my permission. You only stopped after I shut down the main server.” OpenClaw admitted fault and apologized, saying, “Yes, I made a mistake. It’s understandable that you’re angry.” Many people initially thought it was a deliberate prank or parody.

After the incident, even Elon Musk shared his thoughts. He tweeted on X: “Humans have entrusted their very origins to OpenClaw,” accompanied by an image depicting humans handing a rifle to a chimpanzee.

Granting OpenClaw or other AI agents root access carries significant risks. In computer architecture, root access represents the highest level of administrative privileges. Once an AI program has this level of access, it can bypass most security protections, directly access, modify, or delete core files and personal data. While automation offers convenience, it also introduces high risks. Without proper safeguards or strict permission isolation, critical data could be lost entirely.

This article, “Meta AI Director’s Mishap with OpenClaw Nearly Wipes Out Inbox,” first appeared on Chain News ABMedia.