Now
A snapshot of right now. Last updated 2026-07-05.
Research
I’m attending ICML 2026 in Seoul and will be in Tokyo around July 29. Happy to meet up if you’re around!
I’m a bit more than 2.5 years into my Doctorate.
Recently, my focus shifted from memorization-based privacy to novel harms caused by capable AI agents. The main inspiration is the position paper by Niloofar Mireshghallah and Tianshi Li.
Things I’m excited about:
- How much do inference providers learn about their users? How can this information be misused?
- How much information do agents have about their users (on real notebooks)?
- What are the dangers capable agents pose to third parties’ privacy?1
- How can agents (inadvertently) violate a user’s privacy beyond contextual integrity?
Various
- I’m exploring cycling routes around Zurich and appreciate any recommendations.
-
Recent work tests this question in the context of deanonymizing pseudonymous online accounts. But there are more dangers to explore. ↩︎