alt Hacker NewsYou should note that auto-mode is handled by a different session (and likely model) from your actual. There's no road block or deviation or context flooding; all it does is monitor the model actually doing the work and block it from potentially problematic actions (which you can usually override by saying you explicitly give permission for X), nothing more.
The context pollution happens when a denial triggers but the model still keeps attempting the harmful action via alternative avenues. It's got nothing to do with the auto-mode eval model's actual output.