Paper tests iterative LLMs for repairing vulnerable functions
Original title: Found, but Not Fixed : From Software Vulnerability Discovery to Fixes with Iterative LLM Pipelines
The authors compare a multi-stage feedback loop with GPT-4o, GPT-5.1, and Gemini 3 Flash and release weights and data.
The paper argues vulnerability repair, not detection, remains the bottleneck because manual fixes are slow and can introduce regressions. It proposes a generalized multi-stage LLM feedback loop for repairing vulnerable functions in real-world code. The authors say benchmarks against GPT-4o, GPT-5.1, and Gemini 3 Flash show high-quality repair, and they publish the framework, fine-tuned weights, and dataset.
Unconfirmed: The abstract claims high-quality repair but gives no measured success rate.
- Materials are posted at sites.google.com/view/llm4cve.
- The work targets repair of already discovered vulnerable functions.
- No repair success rate or vulnerability set is stated in the abstract.
From the original article
1 min read at arxiv.orgSoftware vulnerabilities persist as a critical threat despite the deployment of sophisticated detection tools and AI assistants. While detection has advanced, the subsequent code repair phase remains a primary bottleneck; manual intervention is inherently slow, costly, and prone to introducing secondary regressions and issues.
By Youssef Gamal · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.