“AI models used for vulnerability discovery are not yet reliable for automated patching because they have a high false positive rate, with Nikesh Arora estimating they might incorrectly patch 30% of items.”