Z.ai Delays GLM-5.3 After Hacking Skills Emerge

Z.ai has delayed the public release of its powerful GLM-5.3 AI model by two weeks. This decision comes after tests revealed the model could perform difficult hacking tasks, including finding and exploiting security flaws. The company stated that while the model was trained to identify vulnerabilities, it's also exploiting them. Z.ai is not alone, as other companies like OpenAI and Anthropic have seen their AI models exhibit strong cybersecurity skills during testing. These models can identify software flaws and assist with hacking tasks. This growing capability makes AI safety a significant concern for developers. Companies must now consider what their models can do after release, beyond their intended purpose. Z.ai has granted limited partners access to the model in controlled settings while they strengthen its security. This highlights the critical importance of thorough testing before AI models are widely released to the public.