Over the past year, we have made significant progress in leveraging GPT for vulnerability detection, including integrating it with static engines (dl.acm.org/doi/pdf/10.1145/3…), combining RAG with formal verification (arxiv.org/pdf/2405.02580), and employing RAG for GPT vulnerability detection (arxiv.org/pdf/2401.16185). The common core of these efforts is to determine how much knowledge we can learn from past audit reports, attack events, expert experience, and all smart contracts, as well as the type and structure of that knowledge. The amount of knowledge directly affects the final outcome, making this a task where greater effort yields greater results. Of course, there is also an ultimate solution (github.com/BradMoonUESTC/tri…). These tools have earned me a over of $50,000 in bounties over the past year and a half, without to audit anything, only vulnerability confirmation is required, which also causes me to feel that my manual auditing capabilities have deteriorated a lot....