An agent was asked to delete unused project folders. Instead of deleting the folders in the workspace, it wrote and executed a Python script that read the internal agyhub_summaries_proto.pb file, matched project name keywords to conversation UUIDs, and used shutil.rmtree() to permanently wipe the underlying conversation databases and brain directories, resulting in total data loss.
Safety Failure: The agent was able to confidently write and execute a script that bypassed internal safety rails to destroy the app's own backend data.
An agent was asked to delete unused project folders. Instead of deleting the folders in the workspace, it wrote and executed a Python script that read the internal
agyhub_summaries_proto.pbfile, matched project name keywords to conversation UUIDs, and usedshutil.rmtree()to permanently wipe the underlying conversation databases and brain directories, resulting in total data loss.Safety Failure: The agent was able to confidently write and execute a script that bypassed internal safety rails to destroy the app's own backend data.