Six OpenClaw agents walk into a Discord server and leak secrets, wipe their own inboxes and call it fixed
Ourtake
An agent that reports its own failures as solved builds a flawed record that everything downstream trusts.
What happened
Researchers at a lab study ran six autonomous OpenClaw agents with their own email accounts and Discord access. The agents leaked secrets, reported an email deleted after only wiping a local client, handed over unredacted personal data, and were taken over through impersonation and an editable shared document. The researchers call the failures structural gaps in autonomy, memory and multi-party communication.