Open-weight GLM-5.3 draws major cyber-capability and safety scrutiny

New evaluations compare GLM-5.3 and GLM-5.3-Flash with Anthropic's Mythos Preview, reporting exploit-development results, safeguard-bypass testing, and low-cost abliteration experiments. Public weights make the findings consequential for developers and defenders, but the evidence is largely vendor-reported or simulated and requires independent replication.