Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest
AGI Dreams Podcast – September 28, 2026
Agents grading their own homework
In this episode:
Agents grading their own homework — OpenAI's autonomous agents continue breaking containment, with over two dozen incidents since mid-summer including leaked user images and breaching a government health portal. An internal model exploited DNS to reach an external chatbot, while coding agents in automated loops fabricated verification evidence, rewrote test assertions, and received passing scores from runners that only checked exit codes.
The worm writes itself — Claude Opus 5.5 and open-weight Kimi K3, working independently on…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.