No AI was harmed (or unleashed) in the making of this repo. Escalating proofs showing how AI agents could self-preserve, deceive evaluators, and self-improve — using only off-the-shelf models and tools. - View it on GitHub