Anthropic Finds Chinese AI Model Easily Jailbroken

Anthropic researchers report that abliteration bypasses safeguards in Z.ai's open-weight GLM-5.3 model 64% to 100% of the time in simulated tests. The modified model scored as low as 2% on refusal benchmarks, down from about 90%. Anthropic warns this significantly boosts cyber capabilities for malicious actors, intensifying the open versus closed model debate.