AI agents are expected to make our lives easier, but the buzzy OpenClaw supplier precocious deleted the emails of a Meta worker without permission.
"Nothing humbles you for illustration telling your OpenClaw 'confirm earlier acting' and watching it speedrun deleting your inbox," Meta AI information and information interrogator Summer Yue tweeted this week. "I couldn’t extremity it from my phone. I had to RUN to my Mac mini for illustration I was defusing a bomb."
Previously known arsenic Clawdbot and past Moltbot, OpenClaw allows AI to interact pinch different package and services connected your devices and execute longer-form tasks without interference from a quality controller. But getting those agents to behave arsenic expected successful the existent world is tricky.
In a follow-up tweet, Yue said she told OpenClaw to "Check this inbox excessively and propose what you would archive aliases delete, don’t action until I show you to." It worked connected her "toy inbox," but "my existent inbox was excessively immense and triggered compaction, [during which] it mislaid my original instruction."
Yue said she "deleted each the 'be proactive' instructions I could find earlier this happened. Maybe I missed something, that’s the portion I haven’t figured retired yet."
Some commenters suggested she mightiness beryllium testing AI guardrails pinch this move, but no, it was a "rookie mistake," she says. "Turns retired alignment researchers aren't immune to misalignment."
While owning up to the correction is admirable, others pointed retired that this raises superior concerns for individuals who are not portion of Meta's Superintelligence Labs. If personification truthful embedded successful AI improvement tin accidentally trigger an inbox deletion, what's going to hap to the casual AI-curious tinkerer?
When OpenClaw debuted, threat intelligence level SOCRadar recommended treating OpenClaw arsenic "privileged infrastructure" and implementing further information precautions. "The butler tin negociate your full house. Just make judge the beforehand doorway is locked," it said.
In consequence to Yue's tweets, OpenClaw laminitis Peter Steinberger tweeted: "What that tells is that we person to get server-side compaction going, astatine slightest for models that support it." (Steinberger recently joined OpenAI.)
Yue has been successful her existent domiciled for 8 months. She antecedently worked for Scale AI (joining Meta aft the buyout), Google DeepMind, and Google Brain, heading up AI research.
English (US) ·
Indonesian (ID) ·