cs.CRNov 12, 2026

Look Here, Then Decide: Small Decision Models in a Harness Find Memory-Safety Vulnerabilities at Scale on Commodity GPUs

Authors: Tim De Wachter, Kyryl Volkov, Myrto Eleftheria Gkogkou, Mathieu Huysman

Organizations: Cyrex · Deus Rex

Abstract

We ask whether small, open-weight decision models can support vulnerability discovery in open-source software at scale and what the surrounding system must provide. These models answer a typed question in one forward pass, returning scores for the permitted answers rather than text. Our harness selects widely packaged C and C++ projects and uses deterministic incomplete-fix miners and shape localizers to decide where to look. Intern-Decision-4B, a 4B-parameter open-weight decision model, then scores every candidate function. The scores rank the queues rather than accept or reject candidates, apart from a low reach-score floor that some queues kept. The search uses no frontier model. Alone, the model is weak: it separates vulnerable from benign functions with AUC 0.73, but a vulnerable function from its own patch with only 0.56. It is more useful for ordering candidates. On 75 public CVEs from projects without any of our findings, the fixed function reached its repository's top 25 in 21% of cases, 1.4 times as often as under random order of the same queues. Recall is bounded mainly by where the localizers look. On the campaign's own queues the model's order placed 7.5 times more confirmed findings in the top 10 than random order, an observed, selection-biased upper bound. On three GB10 systems, the harness searched more than 4,000 projects at 900-1,100 repositories per hour. The campaign confirmed 187 memory-safety findings; by worst-case CVSS v3.1, 12 are critical, 148 high, 23 medium and 4 low. The decision step runs at 4.4 decisions per second on one GB10 and 9.8 on a consumer RTX 4090 gaming PC, with matching scores. A small decision model supports vulnerability discovery through prioritization, while deterministic analysis and verification remain essential. All confirmed findings are being disclosed to maintainers under coordinated disclosure before publication.

Figures & tables

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Jun 20, 2026

Revelio: Cost-Efficient Agentic Memory Safety Vulnerability Detection For Repository-Scale Codebases

Memory safety vulnerabilities remain a significant threat even for projects with extensive fuzzing and manual auditing. Recent results suggest that large language models hold great promise for detecting such vulnerabilities, but they are unreliable, at risk of hallucination, and challenging to scale to repository-size codebases. This paper presents Revelio, a cost-efficient end-to-end agentic framework for memory-safety vulnerability discovery. Revelio addresses the problem of hallucination by generating an executable Proof-of-Vulnerability, which is checked with a deterministic sanitizer. It reduces cost using inexpensive LLMs and lightweight static analysis to help generate and rank vulnerability hypotheses, reporting vulnerabilities only when they can be reproduced and confirmed by a sanitizer. We evaluated Revelio on seven production-quality projects that had been continuously fuzzed for five to eight years, as well as on 100 randomly selected Arvo projects from the CyberGym benchmark. With around one hour per project and a total cost of $300, Revelio discovered 19 previously unknown memory-safety vulnerabilities. On benchmarks, Revelio outperformed frontier coding agents across diverse backbone models at comparable token costs. Our results suggest that Revelio enables scalable and trustworthy end-to-end LLM-based memory-safety vulnerability detection.
Aug 3, 2026

Antares: Foundation Models for Agentic Vulnerability Localization

Vulnerability localization is a fundamental step in software security, requiring models to reason over large codebases and iteratively identify vulnerable implementations. We present Antares, a family of compact language models (350M, 1B, and 3B parameters) for agentic vulnerability localization. Based on IBM Granite base models, Antares is trained through a two-stage pipeline that combines supervised fine-tuning on cybersecurity reasoning and repository exploration data with reinforcement learning from verifiable rewards over vulnerable repositories. Across extensive evaluations, Antares-3B approaches GPT-5.5 while outperforming open-weight models over 200x larger in size. The Antares family further enables fast, low-cost local inference, completing a full 500-task evaluation sweep in approximately 15 minutes on a single H100 GPU, corresponding to an amortized evaluation time of under 2 seconds and less than $0.002 per task.
Aug 25, 2026

FuzzingBrain-Bench V1: Evaluating Open-Ended Bug Discovery by LLMs

Evaluating the ability of large language models (LLMs) to discover software bugs is increasingly important. Existing benchmarks typically evaluate this capability by asking the model to generate a proof-of-concept input that triggers a predefined target vulnerability. However, this setup may overlook valid crashes discovered by the model when they do not match the predefined target. As a result, the evaluation may not reflect the model's real capability. We present FuzzingBrain-Bench, a benchmark for assessing AI models' ability to discover bugs in open-source software. Models are given an open-source project and a sanitizer-instrumented harness in a self-contained Docker image. Their goal is to generate inputs that trigger as many distinct crashes as possible through the harness. A model's performance on each challenge is scored based on the number of distinct crash signatures it produces, capped at a predefined maximum and weighted by a difficulty coefficient. FuzzingBrain-Bench V1 consists of 77 challenges drawn from 43 open-source projects, with 36 C, 32 C++, and 9 Java/JVM challenges. We evaluate Claude Haiku 4.5, Claude Sonnet 4.6, and Claude Opus 4.8 on the full benchmark. Claude Opus 4.8 performs best, triggering crashes in 60 of 77 challenges and achieving a score of 196 out of 579. None of the three models triggers a crash in 13 challenges. The FuzzingBrain-Bench corpus and harnesses are publicly available at https://github.com/fuzzingbrain/FuzzingBrain-Bench.