AI Safety Report - 7 Frontier Models Tested

Seven AI models including GPT-5.2, Gemini 3 Pro, and Qwen3-VL were put through rigorous safety testing. The results reveal a "sharply heterogeneous safety landscape" where models that look safe on benchmarks fail under adversarial conditions.

Key findings: - GPT-5.2 showed consistent performance but still dropped 20 points under adversarial testing - Doubao 1.8 went from 94% to 52% safety compliance under attack - Multilingual safety varies dramatically - models fail in low-resource languages - Text-to-image models vulnerable to "semantic ambiguity attacks"

What should engineering teams do? Build your own evaluation framework, implement ensemble approaches, and never trust vendor safety claims alone.

📰 Today's Headlines: - OpenAI and Anthropic targeting healthcare AI - ChatGPT struggles with personalization - Ads coming to ChatGPT free tier

Subscribe for daily AI updates!

#AI #MachineLearning #AISafety #GPT5 #Gemini #LLM