Kamalika moves to DeepMind

One of the people Meta put on privacy and security for artificial intelligence is now moving to Google DeepMind to work on Gemini instead. Kamalika Chaudhuri said she is joining Google DeepMind to lead work on Gemini’s security and privacy after serving as a director and research scientist at Fundamental Artificial Intelligence Research, Meta’s research arm. (x.com, cseweb.ucsd.edu) That is a very specific job at a very specific moment. Google DeepMind says its responsibility and safety work includes “privacy preserving infrastructure and models” that are adapted into Gemini and products, which means this is not a generic research hire. (deepmind.google) Gemini is the family of models Google uses in its chatbot, developer tools, and other products, so “security and privacy” here means protecting both the model and the data the model can touch. Google DeepMind has already published work on Gemini defenses against indirect prompt injection, a trick where hidden text in an email or document tries to make the model leak data or misuse permissions. (deepmind.google) Chaudhuri’s background fits that job unusually well. On her University of California, San Diego page, she says her work covers artificial intelligence privacy, security, safety, and alignment, including measuring privacy and security leaks in large language models. (cseweb.ucsd.edu) She is also one of the better-known researchers in differential privacy, which is a way of training or analyzing data so the system learns patterns about the crowd without exposing too much about any one person. Her publications include early papers on differentially private machine learning, private empirical risk minimization, and private stochastic gradient descent. (cseweb.ucsd.edu, cseweb.ucsd.edu) That matters because modern assistants are no longer sealed-off chat boxes. Google’s own safety material describes models that can read documents, calendars, and websites, which turns privacy from an abstract policy issue into an engineering problem about what the model can see, remember, and reveal. (deepmind.google, ai.google.dev) Meta had been building up that same muscle from the other side. Chaudhuri’s page says she led a FAIR team working on artificial intelligence privacy, security, and reliability, while Meta says it has spent more than $8 billion on its broader privacy program since 2019. (cseweb.ucsd.edu, about.fb.com) Google DeepMind has been formalizing this work too. Its updated Frontier Safety Framework says the company has implemented the framework in safety and governance processes for frontier models including Gemini 2.0, and Google’s 2026 Responsible Artificial Intelligence Progress Report lists dedicated leadership roles for trust, safety, and responsibility. (deepmind.google, blog.google) So this hire is less about one résumé line than about where the frontier labs are spending senior attention. A few years ago, the prestige hires were mostly model architects and scaling researchers; now the same labs are visibly recruiting people whose specialty is stopping models from leaking secrets, following malicious instructions, or mishandling sensitive data. (cseweb.ucsd.edu, deepmind.google, deepmind.google)