LLM internal states reveal code vulnerabilities, study finds

Researchers have developed a method to detect code vulnerabilities by analyzing the internal activations of large language models (LLMs) rather than just their final output. By training small probes on the latent activations of models like Granite 4.1 8B and Qwen3.5-9B, they found that these internal states can signal the presence of vulnerabilities in C/C++ code. While the probes achieved competitive results on some benchmarks, they still lagged behind state-of-the-art classifiers on more challenging datasets, indicating potential for future development in model-native vulnerability screening. AI

IMPACT This research suggests a new avenue for lightweight, model-native vulnerability screening in AI-generated code.

RANK_REASON Academic paper detailing a novel method for code vulnerability detection using LLM internal states. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Big-Vul
• Devign
• Draper VDISC
• Gemma 4-12B
• Granite 4.1 8B
• PrimeVul
• Qwen3.5:9b
• Qwen3.6-27B

• paper
• safety
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →