EVENT DOSSIER
The researchers proposed the BUGSTONE-E2E framework, aimed at converting fixes submitted in the public vulnerability database into executable detection rules and verifying them. This framework uses Tree-sitter to enumerate and match anchor sites, combining lightweight heuristic methods to eliminate benign sites. Subsequently, agents based on large language models (LLMs) check the remaining candidates under the guidance of the rules, ultimately constructing runtime verifications and generating patches verified by bidirectional differential testing. Based on 19,325 high-risk CVEs from 2022 to 2026, the experiment successfully identified 2,710 fixes, constructing 1,033 detection rules covering 56 CWE families (encapsulated into 172 skills), and generating 644 runtime evidence of findings in 14 programs.
研究人员提出 BUGSTONE-E2E 框架，将公共漏洞数据库中的修复提交转化为可执行检测规则并进行验证。该框架利用 Tree-sitter 枚举匹配锚点的调用站点，结合轻量级启发式方法剔除良性站点，随后由基于 LLM 的代理在规则引导下检查剩余候选项，最终构建运行时验证并生成经双向差分测试校验的补丁。实验基于 2022 至 2026 年 19,325 个高危 CVE，识别出 2,710 个修复提交并构建了涵盖 56 个 CWE 家族的 1,033 条检测规则（封装为 172 项技能），在 14 个程序中产生了 644 个发现的运行时证据。