Research

Exploit-guided repair in which the repair objective adapts. An exploitation agent re-attacks every accepted patch, and each execution-confirmed bypass becomes a permanent requirement the final patch must block. The exploits used during construction are kept strictly separate from held-out evaluation suites. On a 101-CVE Java and C/C++ benchmark, held-out fix-closed coverage rose from 78.1% to 84.7%, and beyond-upstream coverage from 42.1% to 62.5%, against single-pass repair on the same agent and model.

A multi-stage LLM pipeline that takes raw static-analyzer alerts and triages them through contextual reasoning and structured evidence validation, reducing false positives without sacrificing recall. Evaluated 10 frontier LLMs across 6 model families on the OWASP Java Benchmark (1,974 cases, 10 CWE categories) and CWE-bench — 755 real CodeQL alerts across 56 project–CVE pairs from 37 open-source Java repositories. CWE-specialized prompting improved F1 by up to +0.26 on real-world code; best F1 is 0.912 on OWASP and 0.837 on CWE-bench.