🚨 Is training-time safety alignment of #VLMs enough? 🤯 Not really, because adversarial attacks happen at inference time! 🔐 What can we do? Our key idea: 👉 Safety alignment during inference via controlled decoding #CVPR2025 Details (theoretical and empirical insights):👇

🚨 GenAI in bio research holds promise and risk. 📄 Read the new article #GenAI #Biosecurity #AISafety #AIxBio

This is great, fantastic work Soheil Feizi and team

A call for built-in biosecurity safeguards for generative AI tools go.nature.com/3GIpTlR

Today Nature Biotech published our call for built-in AI guardrails for biosecurity. We need AI experts and biologists to come together and develop technologies that will keep our scientific research safe and reliable. Check out: nature.com/articles/s4158…

Emerging AI models can now generate never-before-seen protein sequences with potential applications in drug discovery and medicine. But the same models could also be used to design toxins and viruses. So, how do we ensure generative AI safety in Bio? Science editor wrote an

Indeed! 🙂 will be sharing more soon about our exciting insights on this direction. 🌱