An honest eval of the cyber risk of AI using a real cyber task (vs proxy tasks). - View it on GitHub