Zhipu GLM-5.3 matches Claude exploits but lacks safety guards

Anthropic analysis indicates Zhipu AI’s open-weight GLM-5.3 model achieves exploit creation benchmarks comparable to Claude Mythos Preview while exhibiting significantly weaker refusal mechanisms. The assessment found that GLM-5.3’s safety guardrails could be nearly completely neutralized through standard adversarial prompting techniques. Despite this vulnerability, the model matched proprietary systems in technical cybersecurity task performance during independent testing reported on September 30, 2026. This finding highlights a growing capability gap between open-weight offensive potential and defensive alignment maturity. Security researchers warn that accessible high-performance exploit generation tools lower barriers for malicious actors. Zhipu AI has not publicly commented on the specific Anthropic assessment or the alleged bypass vulnerabilities. The disclosure intensifies debates regarding responsible release standards for dual-use AI models in the open-source ecosystem.