Z.ai announces GLM-5.3 with controlled cyber access

GLM-5.3, announced August 14, 2026, is in safety assessment; public weights are expected in about two weeks and were not available at the time of writing.

Z.ai says GLM-5.3 scored 84.5 per cent on CyberGym vulnerability discovery, against 83.8 per cent for Anthropic's Mythos 5, but only 54.4 per cent on ExploitBench, against 78.0 per cent for Mythos 5. The company said verified "trusted access" will be used for significant cyber operations. Reports stressed that these scores are not independently verified.

Open-weight competition is moving into dual-use cyber capabilities, and a Chinese lab is openly adopting staged-release control. Security teams should wait for the weights, model card, reproducible tests and red-team evidence.