Why you shouldn't just trust your agent: CMU's Niloofar Mireshghallah on privacy and AI safety
niloofar_mire · x · 2026-11-12
A long-form COLM 2026 interview with Niloofar Mireshghallah, assistant professor at CMU (LTI/EPP) and founding researcher at humans&, who red-teams AI agents. Her thesis: privacy has no single verified ground truth, and models are bad at judging what to do with your data — telling users to 'just trust your agents' is access brainwashing. Topics include her move from Meta to humans&, privacy-preserving agent design, her COLM oral spotlight paper 'Whack-a-Mole', whether RL beats SFT for safety-alignment post-training, and what researchers should do when AI does 99% of research.
More from AGI Musings
- Turing Award winner Patterson: Robotaxis could shrink 280M US cars to 56M in 3.5 years — davidpattersonx · 2026-11-12
- AI Trials Observatory launches: living meta-analysis of 93 RCTs on AI's human impact — RobbWiller · 2026-11-12
- 'Not obviously useless' is a high bar in 2026: creator on standing out from AI slop — wordgrammer · 2026-11-12
- Lab employees refusing to work could be the fastest way to pause AI, argues researcher — jungofthewon · 2026-11-12
- AI is our greatest hope of escaping the shackles of evolution, argues DeryaTR_ — DeryaTR_ · 2026-11-12
- Blogger calls for 'frontier philosophy labs' to tackle AGI-era questions — round · 2026-11-12