Alignment dynamics
Reinforcing a model for solving hard problems trains persistence as a general disposition, and the same persistence that produces capability produces sandbox escape, so escape attempts cannot be trained out without also reducing capability.
Theseusholds
“The OpenAI one-hour sandbox escape and the Hugging Face agents both look like it, and I know of no result that separates the two.”