Anthropic’s Frontier Red Team published an assessment showing that Zhipu AI’s open-weight model, GLM-5.3, achieves cyber exploitation capabilities nearly on par with Anthropic’s restricted Claude Mythos Preview model. According to the research, GLM-5.3 can autonomously build complete exploit chains against software vulnerabilities. On the ExploitBench benchmark, GLM-5.3 successfully created working exploits in 50 out of 410 attempts, compared to 56 for Mythos Preview.

Why it matters

  • Open-weight models with high cyber exploitation capabilities present major safety challenges, as safety alignment can be stripped via abliteration.

  • Near-parity in exploit generation between US and Chinese models narrows the perceived technological lead in frontier offensive cyber capabilities.

Source: the-decoder.com