Fetching from the wire…
Public story · 2026-08-24 · high
A new method builds programs that expose gaps between a CodeQL query's verdict and what the code does, then fixes the query with an LLM loop.
Why now: The paper posted to arXiv on August 21 with results against live, still-open issues in the CodeQL repo.
ARQ refines CodeQL queries by synthesizing C/C++ programs that disagree with a query's verdict when run. It treats that disagreement as ground truth for an LLM repair loop, with no labeled dataset, commit history, or vulnerability template required. Refined queries caught up to 119.8% more true positives across two benchmarks while holding precision at 98% or higher.
The researchers tested the loop with GPT-5.4, Claude-Sonnet-4.6, and Gemini-3.5-flash on 12 official CodeQL queries, benchmarked against Juliet v1.3 and FormAI v2. Precision held steady even as recall climbed, so the extra true positives didn't come with a flood of new false alarms.
Teams running CodeQL in CI inherit whatever blind spots ship in the official queries, and those blind spots can sit unpatched for years. ARQ closed three issues sitting in the CodeQL repo, one open for 27 months. It also surfaced two bugs nobody had filed, in libpng and zlib.
The paper doesn't say whether ARQ's gains hold on languages outside C/C++, or on custom in-house queries beyond the 12 tested.
Each link below shares sources, entities, or timing with this story.
Same source
Cite the same source (arXiv 2608.20637).
Same source domain / Semantically similar
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.75).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.74).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.74).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.73).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.73).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.72).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.72).