GLM 4.7: Open-Source AI Matches Proprietary Safety

Open-source AI models can now match or even surpass proprietary models in safety, according to a new study from TELUS Digital. This challenges the idea that open AI is inherently riskier for businesses. The benchmark involved over 620,000 tests on 34 AI models from 10 global providers. What's interesting is that safety appears to depend more on a model's design, size, and reasoning ability, rather than whether it's open-source or proprietary. For example, the open-source model GLM 4.7 outperformed several proprietary competitors. No AI model was found to be completely immune to attacks, with vulnerability rates ranging widely from 1.3% to 93%. Smaller AI models were generally the most vulnerable. Reasoning models were significantly harder to exploit, showing a 19.9% vulnerability rate compared to 55.1% for non-reasoning models. The study also highlights a significant underinvestment in AI security by enterprises. This research matters because it could change how organizations approach AI adoption and security strategies.