Put in simpler terms, when a frontier LLM generates a vulnerability patch autonomously, there is only a roughly 1 in 4 chance that it will do so successfully. There is a roughly 50-50 chance that it will fail to fix the original bug, a 1 in 4 chance that it will introduce a new bug in general, and nearly a 1 in 20 chance that it will introduce a new security vulnerability specifically. As such, the expected value of a fully LLM-generated, non-human-reviewed patch is a net-negative by a considerable margin.

Just a few more tokens, trust me bro.