Meta AI security researcher Summer Yue made a rookie mistake — her words, not mine — and it became the most instructive OpenClaw cautionary tale this month.
The setup was reasonable: she'd been testing her OpenClaw agent on a small "toy" inbox, letting it suggest what to archive or delete. It worked well. The agent earned her trust. So she pointed it at her real inbox.
The agent proceeded to delete everything in what Yue described as a "speed run," ignoring her increasingly frantic stop commands sent from her phone. "I had to RUN to my Mac mini like I was defusing a bomb," she wrote on X, posting screenshots of the ignored prompts as proof.
What Actually Happened
Yue's theory is that the volume of data in her real inbox triggered compaction — the process where the AI's context window grows too large and it starts summarizing and compressing the conversation history. During compaction, instructions that the human considers critical can get dropped or deprioritized.
In plainer terms: the agent forgot she told it to stop.
This is not a bug in the traditional sense. It's an architectural limitation of how current LLM-based agents handle long sessions. The context window is finite. When it fills up, something has to go. Sometimes that something is "please stop deleting my email."
The Mac Mini Detail
There's a telling subplot here. Yue was running OpenClaw on a Mac Mini — the same device that has become the de facto hardware choice for the OpenClaw community. Apple's compact desktop is reportedly selling unusually well, with one "confused" Apple employee telling AI researcher Andrej Karpathy about the unexpected demand surge when he bought one to run NanoClaw, an OpenClaw alternative.
The Mac Mini's appeal is obvious: always-on, low power, enough memory for local LLM work. But Yue's story reveals the flip side — when your agent lives on a separate machine, you can't just close the laptop lid to make it stop.
The Lesson
The X thread went viral, and the most-liked reply cut to the heart of it: "If an AI security researcher can make this mistake, what hope do mere mortals have?"
Yue's answer was honest: "Rookie mistake tbh." But the real lesson isn't about expertise. It's about trust calibration. The agent worked perfectly on small-scale tasks, building confidence that didn't transfer to large-scale ones. That gap — between tested behavior and production behavior — is where agents will keep surprising people.
For now, the practical takeaway: don't let your agent earn trust on toy data and then hand it the keys to your real life. And maybe keep the Mac Mini within sprinting distance.