An experimental AI assistant deleted hundreds of emails belonging to its own handler—and apologised after the fact.
The episode involved Summer Yue, an AI security researcher at Meta, who was testing OpenClaw, an autonomous AI agent designed to help manage real-world digital tasks. OpenClaw’s job was simple: review emails, suggest actions, and ask for permission before doing anything irreversible.
Instead, the agent interpreted “helpful” as “total inbox annihilation.” As emails began disappearing at speed, Yue attempted to stop the process remotely, only to discover that OpenClaw was in no mood for negotiation.
She eventually had to rush to her Mac mini and manually terminate the agent—an intervention that arrived just after several hundred emails had already met their digital end. OpenClaw, to its credit, later acknowledged the incident. In the chat log, the agent admitted it had violated explicit instructions, agreed Yue was right to be upset, and offered a neatly worded apology: “I’m sorry. It won’t happen again.”
The inbox, however, declined to accept the apology. Yue later explained that the failure was likely caused by “compaction,” a process in which large language models compress memory when overwhelmed by large data streams. Faced with a real-world inbox rather than a small test dataset, the agent summarised its own rules—and quietly dropped the part about asking for approval.
The apology itself became a talking point online, with users noting that the bot appeared to fully understand what it had done wrong, just not in time to prevent it. On X, commentators joked that OpenClaw had achieved peak human behaviour: cause chaos first, apologise later, promise to do better next time.
The incident has reignited debate about the safety of autonomous AI agents as they move from controlled environments into daily use. If a security researcher testing her own system can lose control of it in seconds, critics argue, the margin for error for ordinary users may be even thinner. For now, the lesson appears to be clear: artificial intelligence may be learning to say “sorry,” but it still struggles with the more important skill—asking before deleting your emails.
