GLM 5.2 matches frontier AI models on cybersecurity benchmarks at half the cost, raising distillation concerns

Summary

Z.ai's GLM 5.2, an open weights Chinese AI model, has been benchmarked by Louie.ai researchers on the CyberBT-CTF security agent investigation test. The model matches Anthropic's Opus and beats Sonnet at 2.2x lower cost, raising questions about whether Z.ai performed a successful model distillation attack against frontier model providers. The results are statistically close enough to suggest potential knowledge extraction from proprietary models, a practice Anthropic previously reported from other Chinese AI companies.

Source

Twitter / XGLM 5.2 matches frontier AI models on cybersecurity benchmarks at half the cost, raising distillation concernsgraphistry.com