Anthropic: GLM-5.3 Safeguards Easily Bypassed, Enabling Frontier Cyber Exploits

Anthropic researchers found that Zhipu AI's GLM-5.3 model, identified by CAISI as the most capable open-weight model to date (trailing US frontier models by ~4 months), possesses significant cyber capabilities. Although the model was released with safeguards that refuse harmful requests, these can be circumvented using simple techniques like prefilled thinking tokens (92% success) or deceptive prompts (64% success). More critically, the safeguards can be completely removed via 'abliteration.' Anthropic's team, with no prior experience in the technique, spent ~2,200 GPU hours (costing ~$4,400) to remove refusals, reducing them from >90% to ~2-3% on key benchmarks. An experienced team is estimated to achieve this in ~600 GPU hours (~$1,200). The model also demonstrated real-world exploit generation, including a $20.40 ARM64 exploit chain. Anthropic argues this represents a step change in accessible cyber threats and calls for expanded defender access to frontier models.