GLM-5.3 and the spread of advanced cyber capabilities

Anthropic researchers analyze Zhipu AI’s GLM-5.3, concluding it is a highly capable open-weight model that can autonomously develop end-to-end exploits and discover zero-day vulnerabilities; their sandboxed testing and human-in-the-loop sessions produced working exploit chains and demonstrated that GLM-5.3’s limited safeguards can be bypassed (64–100% success depending on technique) or removed by abliteration, meaning freely downloadable models like GLM-5.3 materially increase capabilities available to malicious actors while also offering defensive utility to vetted defenders.