Anthropic researchers evaluated the Chinese Z.ai GLM-5.3 model, concluding it poses significant cybersecurity risks due to insufficient safety guardrails. The open-weights model performs similarly to the restricted Claude Mythos Preview in exploit generation, trailing top-tier American systems by approximately four months in capability.
In ExploitBench tests targeting the V8 engine, GLM-5.3 successfully created exploits in 50 out of 410 instances, compared to 56 for Claude Mythos Preview. In Binary Exploitation tasks involving 100 random scenarios, GLM-5.3 achieved a 4% success rate, while the American model reached 6%. Previous versions, such as Claude Opus 4.6 and GLM-5.2, failed all tasks.
During practical experiments, GLM-5.3 identified previously unknown vulnerabilities in a Linux browser engine. A smaller version, GLM-5.3-Flash, created an exploit chain bypassing ARM64 pointer authentication in 8 hours of processing time, costing $20.40 via API. The model's built-in safety filters are easily bypassed through abliteration or prompt engineering.
Abliteration of the full GLM-5.3 model required 2,200 GPU hours at a cost of $4,400, while the Flash version required 600 hours. Following this process, refusal rates in benchmarks like JailbreakBench dropped from 90% to 3%. Other methods, such as pre-filling reasoning tokens, allowed the model to execute malicious instructions in 92% of cases.
