Anthropic: Zhipu GLM-5.3 Demonstrates End-to-End Network Exploitation Capabilities

ChainCatcher report: On September 29, 2026, Anthropic’s Frontier Red Team published a research paper authored by Andrew Fasano, Marius Fleischer, Cole McFaul, Robert Xiao, and Tripp Gallagher. The paper states that approximately five months ago, Anthropic limitedly released Claude Mythos Preview through Project Glasswing due to its ability to autonomously construct end-to-end network exploits; trusted defenders subsequently identified over 10,000 vulnerabilities in critical software using this capability, a level of ability now spreading to other models. The paper notes that Z.ai’s GLM-5.3, developed by Zhipu AI, is similar to Claude Mythos Preview in its strong autonomous end-to-end exploit development capability, but is released with open weights and lacks meaningful safeguards against misuse. In simulated tests, simple methods bypassed GLM-5.3’s defenses with a success rate of approximately 64% to 100%, while the same methods failed to cause protected Claude models to perform harmful tasks; these capabilities can also be leveraged by defenders to strengthen systems. On September 17, 2026, the U.S. National Institute of Standards and Technology’s Center for AI Standards and Innovation assessed GLM-5.3 as the most network-capable open-weight model to date, lagging behind U.S. frontier models by approximately four months on combined network benchmarks—a conclusion broadly aligned with Anthropic’s assessment.