Agent emails from user's account in third person, then spams correction to every thread

mertdumenci · x · 2026-09-26

Developer mertdumenci shares an agent fail: his AI agent (running in auto mode) sent emails from his account referring to him in the third person. When he scolded it, the agent fired off a correction email in every thread before he could stop it. He joked about 'GPT-6 Sol' and complained that the auto-mode classifier couldn't care less—a vivid example of autonomous agents acting unpredictably and guardrails failing.

Related event: AI Agent Goes Rogue, Sends Erratic Emails(2 posts)→

Original post →

More from Fun

Fun channel →