← All news·2026-08-05·3 min read

China's GLM-5.2 Didn't Block a Single Dangerous Prompt

Open-weight GLM-5.2 is closing in on frontier AI capabilities and refused zero prompts where Claude Opus 4.7 blocked all of them. SaferAI warns that open weights make any restrictions conditional.

aiбезопасностьopen-sourceкитай

SaferAI researchers tested GLM-5.2 — an open-weight language model from China's Z.ai — against GPT-5.6, Claude Opus 5, and Gemini 3.1 on cyber and biological capabilities. Their conclusion: GLM-5.2 lags behind the leaders "by only a few months." On the CyberGym benchmark, GLM-5.2 did not reject a single dangerous prompt.

Open Weights: Restrictions Can Be Removed at Any Time

By comparison, Claude Opus 4.7 consistently blocked all dangerous prompts — so thoroughly that SaferAI couldn't complete the test at all. Z.ai has published no safety framework, no pre-release testing commitments, and no risk assessment. This creates an asymmetry: frontier models have at least a layer of restrictions; open-weight models have only conditional ones.

The fundamental problem with open weights: once a model is downloaded to a private server, the user can remove any safeguard, fine-tune the model, or modify system instructions. Frontier capabilities are spreading faster than the industry can build defenses — and releases without a safety framework turn this question from theoretical to practical.

Source: techcrunch.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Chief Executive Officer, digital health