News · AI

Grenoble study achieves 60.0% code vulnerability detection in one second with LLM probing

French researchers introduced REVA, extracting contextual call-graphs and threat models in a single forward pass to match agentic detectors while running 250 times faster.

60.0%

accuracy in real-time LLM vulnerability detection on PrimeVul

Grade 2 · Published method

Source: arXiv

What changed
Researchers at Université Grenoble Alpes introduced REVA, achieving 60.0% vulnerability detection accuracy in about one second per function on the PrimeVul benchmark.
Who it counts
The evaluation covers C and C++ security vulnerability detection across PrimeVul's held-out test suite using GLM-5.3-Flash and two additional open language models.
What it does not tell you
The study benchmarks open-source repository functions and does not test performance on proprietary enterprise monolithic codebases or non-C languages.

The full story Check the source

Researchers at Université Grenoble Alpes developed an artificial intelligence security tool that achieves 60.0% accuracy in detecting software vulnerabilities in approximately one second per function on the PrimeVul benchmark [1].

The system, named REVA, extracts contextual threat models, lightweight static analysis findings, and call-graph data into a single forward pass [1]. This matches the 59.2% accuracy of state-of-the-art agentic detector VulTriage on the same GLM-5.3-Flash model, while operating 250 times faster and at less than one-tenth the compute cost [1]. On the same prompt, the trained internal probe proved 14% to 17% more accurate than the language model's direct generated verdict [1].

What it measures

The evaluation measures vulnerability detection accuracy across PrimeVul's held-out test suite of C and C++ functions [1]. Latency and accuracy were benchmarked across three open language models (GLM-5.3-Flash, Qwen-2.5-Coder-7B, and Llama-3.1-8B) on standardized compute hardware [1].

What it does not tell you

The study evaluates function-level vulnerability detection in open-source projects and does not measure end-to-end exploit synthesis or false positive rates in live software development pipelines [1]. The benchmark dataset consists of C and C++ functions, leaving untested how the representation probe performs on memory-safe languages such as Rust, Go, or Python [1].

Sources and their grades

  1. arXiv, Reva: Real-Time Vulnerability Detection through Contextual Probing of Large Language Models. Read 27 October 2026. Grade 2 · Published method arXiv:2610.35829v1 [cs.CR, cs.AI], submitted 2026-10-26 17:50:09 UTC by Amine Lbath and Massih-Reza Amini (LIG / Université Grenoble Alpes, France). Stage 2 computer science preprint released on arXiv; not peer-reviewed.

A grade says how much weight a source will take, from 1 (a body that can compel an answer and has nothing to sell) to 5 (somebody describing what they sell, or themselves). The standard sets out what each grade may be used for.

Filed under AI research, Cybersecurity, Software engineering

Comments

Commenting needs an account, made with nothing but an email address. A first comment is read before it appears.

More news

See all