We decide to not control the harness, that’s what resulted in the HF hack where OpenAI decided to run thousands of agents in parallel, using a model that was tr

We decide to not control the harness, that’s what resulted in the HF hack where OpenAI decided to run thousands of agents in parallel, using a model that was trained for attack, with a harness that allows full execution, with close to no supervision. For months. But sure, people can run models with a different harness, though in that case it’s not really escaping anything. The harness is deterministically doing something with negative impact