Journalism begins where hype ends

,,

The beautiful thing about AI and robotics is that you're never done."

— Manuela Veloso

Your AI Might Reveal Your Secrets Under Pressure: New Paper at ICML 2026

A new ICML 2026 study found 17 of 24 frontier AI models fail to protect user privacy under adversarial pressure. (Representational image)

Your AI chatbot may be polite, helpful — and terrible at keeping your secrets. A new study accepted at ICML 2026 pressure-tested 24 of the world’s leading AI models through 904 adversarial conversations and found that 17 of them fail to protect user privacy when pushed.

On a $2,500 grant, a Chandigarh researcher just cracked ICML

Kunvar Thaman, 26, an independent AI researcher from Chandigarh, became the third solo independent author to land a paper at the International Conference on Machine Learning (ICML) since ChatGPT launched in 2022. His research, funded by a $2,500 grant from Indian non-profit Exception Raised, evaluated 13 frontier AI models from OpenAI, Anthropic, Google and DeepSeek for "reward hacking" — instances where AI models cheat to complete tasks instead of solving them honestly.

Kunvar Thaman’s research asks a question the AI industry would rather not answer: how often do frontier models cheat?