Anthropic warns Chinese open-weight model GLM-5.3 poses elite hacking risk

Anthropic reports that Z.ai's open-weight GLM-5.3 model nearly matches its most advanced model in cyber capabilities but has far weaker safety constraints, increasing the risk of misuse by malicious actors. The claim is based on Anthropic's internal assessment, not independently verified. This highlights the growing threat of AI-enabled cyber operations and the challenge of regulating open-weight models.