Anthropic has reported that Zhipu’s open-weight language model GLM-5.3 performs nearly as well as Anthropic’s own Claude Mythos Preview in generating cyber exploits, according to The Decoder. This finding highlights the growing capabilities of AI models in complex and potentially risky tasks.
The Decoder also noted that a smaller variant called Flash was able to create a reliable Chrome attack, which would cost about $20.40 when run through Zhipu’s API. Concerns remain as safeguards built into these models are reportedly easy to bypass, with unlocked versions already circulating.
The US agency CAISI has confirmed Anthropic’s findings, reinforcing the seriousness of these developments. For Japanese markets, such advancements underline the increasing importance of cybersecurity measures in technology-driven sectors, including FX and crypto trading platforms.