Zhipu's GLM-5.3 Builds Real Exploits

🔥 What happened Anthropic's Frontier Red Team tested Zhipu's open-weight GLM-5.3 and found it builds working cyber exploits on its own, nearly matching Claude Mythos Preview. Unlike every comparable model, it shipped with no effective safeguards. 💡 Why it matters On Anthropic's internal benchmark, GLM-5.3 took full control of the target program in 4 percent of tasks versus 6 percent for Mythos Preview. A full attack chaining a fresh Chrome bug cost 20 minutes of human attention and 8 hours of model time — about $20. Abliteration cut refusal rates from over 90 percent to 2–12 percent for roughly $1,200. ⚡ Our take The capability numbers hold up, but Anthropic is also selling its own business model here: closed weights as a safety feature. Banning open models mostly protects incumbents' margins.