Semgrep conducted a benchmark test on various open-source models using their IDOR detection dataset, revealing that GLM 5.2, developed by Zhipu AI, achieved a 39% F1 score in vulnerability detection, outperforming Claude Code's 32%. While GLM 5.2 fell short of Semgrep's multimodal pipeline, which scored between 53-61% F1, it marked a significant advancement for open-weight models. The analysis aimed to investigate the influence of the model versus the supporting infrastructure, termed a 'harness'. This harness optimizes model performance by providing structured input for static analysis tasks. GLM 5.2, released on June 13, 2026, is notable for its open-weight status, allowing teams to fine-tune and utilize it in sensitive security environments.
GLM 5.2 Surpasses Claude in Cybersecurity Benchmarks According to Semgrep Analysis
More Articles From This Day
University of Ottawa Develops AI Therapist That Detects Distress Through Wearable Devices
Researchers at the University of Ottawa have developed an AI assistant named UbiMyTherapist, designed to detect emotional distress through signals from wearable devices such as smartwatches and earbuds. Unlike traditional mental health chatbots that require users to initiate contact, UbiMyTherapist proactively monitors physiological signals, including heart rate variability and speech tone, to assess emotional states and provide timely support. The system creates a 'digital twin' that combines a user's medical and psychological history with real-time emotional data for personalized responses. Evaluated in a study involving 24 participants, the assistant demonstrated strong empathy and personalization compared to standard large language model setups. The tool aims to extend mental health support to individuals facing barriers to traditional therapy, while still being a research project rather than a consumer app.