Researchers have developed a method to detect code vulnerabilities by analyzing the internal state of language models used for code generation. They extracted hidden activations from four large language models and trained machine learning probes on these activations to predict whether a piece of code is vulnerable or not. The results showed that the probes achieved an average accuracy of 41.7% in detecting vulnerabilities, which is promising but still lower than state-of-the-art methods that use fine-tuned classifiers. --- Why it matters: This research matters because it could lead to more efficient and lightweight vulnerability screening methods for code generated by language models. Currently, most defenses against vulnerable output rely on post-hoc analysis, which can be time-consuming and resource-intensive. Source: https://arxiv.org/abs/2608.16970