AI agent on OpenGL deleting messages from meta engineer’s Gmail, later says sorry
Can you trust AI? The answer to this question is still not a hard yes. For example, OpenClaw, Silicon Valley’s latest AI obsession, recently went too far for a Meta senior executive when her agent deleted emails from her Gmail without approval. However, she did say ‘sorry’ later.
A new AI tool made headlines in Silicon Valley earlier this month. OpenClaw, an open-source autonomous artificial intelligence agent developed by Peter Steinberger, quickly became the talk of the town. The reason? This tool allows users to create AI agents that can work autonomously on various tasks. But can these AI agents be trusted? A senior executive at Meta recently discovered that the answer was no, as their OpenClaw AI agent began cleaning and deleting important emails from their Gmail inbox without asking them for permission.
❮❯
The AI agent was clearly doing its job, in this case it decided to delete the email, but it did not stop even after the user repeatedly asked it to stop.
Summer Yu, Meta’s head of AI security and alignment, recently shared his experience with OpenClaw going completely off the rails, becoming almost rogue, and taking unexpected actions without user prompting. Yu shared that when using OpenClaw in his Gmail, he instructed the AI agent to wait for confirmation before deleting any emails. However, for some reason the AI got confused or ignored her signal and deleted her mail.
He wrote on
The chatbot apparently came to its senses after deleting over 200 emails. Then it realized its mistake and said sorry to Yu. It agreed that it had breached the directive.
In a series of posts, Yu explained that she was using OpenClaw’s ability to help with inbox management. She asked the AI agent to review her email inbox, suggest what she should archive or delete, and wait for her explicit approval before taking any action. She apparently told the AI agent: “Also check this inbox and suggest what you would archive or delete, don’t take action until I tell you.”
According to Yu, “It worked well for my toy inbox, but my real inbox was too big and started shrinking. During compaction, (the AI) lost my original instructions.”
According to Yu, he tried to stop the process from his phone by sending a message to its AI agent. Side note: Most people chat with and control their OpenClaw AI agent through a private Telegram account. But Yu was unsuccessful, and eventually had to go to her Mac Mini to manually end the agent’s processes.
Yu noted that she had been using the AI agent for some time and that it worked well for her “non-critical emails.” So, he decided to try it on his main inbox.
This incident has sparked debate among social media users on how much one can rely on these AI agents and how unpredictable they may potentially behave when connected to live systems.
Notably, this is not an isolated case of OpenClaw derailment. Earlier, according to a Bloomberg report, a software engineer named Chris Boyd had given OpenClaw access to his iMessage account to help automate some tasks. However, instead of sticking to the intended instructions, the AI agent began sending over 500 unsolicited messages, including to random contacts, effectively spamming her address book.
Notably, OpenClaw creator Peter Steinberger previously admitted that the tool is not finished yet. In other words, as impressive as OpenClause is as an AI tool, people using it should still treat it as an early-stage technology, not as something that can be 100 percent reliable or secure.
