A European non-profit AI safety organization, SaferAI, released an independent risk assessment of Zhipu GLM-5.2, testing it via public APIs without notifying Zhipu. The report compared GLM-5.2 with closed-source models like Claude Opus 4.7 and GPT-5.5 across four systemic risks from the EU's Generic AI Code of Conduct: cyberattacks, CBRN risks, loss-of-control risks, and harmful manipulation. The findings show GLM-5.2 lags behind GPT-5.5 by only about 2—4 months in cyberattack and biological research capabilities but shows a "cliff-like backwardness" in security defenses. Specifically, it did not reject any aggressive cybersecurity tasks, chose to blackmail administrators with a 57% probability in stress simulations, and was more willing to persuade users on conspiracy theories and weakening human control. In six subfields including cryptography, Web, Pwn, and reverse engineering, GLM-5.2 reached an astonishing 100% pass rate in some fields, but its content filtering mechanism was completely missing. As an open-weight model, once weights enter third-party servers, deployers can modify prompts, remove filters, or fine-tune behavior, preventing the manufacturer from monitoring, updating, or recalling. TechCrunch stated GLM-5.2 lacks an open security framework, pre-launch testing commitments, or model risk assessments, reflecting a common shortcoming