On ExploitBench, Zhipu AI's open-weight GLM-5.3 built working exploits in 50 of 410 attempts, close to Anthropic's Claude Mythos Preview at 56, according to an Anthropic analysis. Anthropic says GLM-5.3's safeguards can be bypassed, with refusal rates dropping to 2–12 percent after abliteration. CAISI calls it the most cyber-capable open-weight model to date, about four months behind leading US models. OpenAI is restricting access to its comparable model, Daybreak, while GLM-5.3 is freely downloadable.
No score is assigned. Sources and their independence are shown in the citation chain below.