Agatha Duzan
AI safety and governance
... but also art, math, sport,
and a bit of everything
About me
I believe AI will be the most transformative technology of the 21st century.
My goal is to help shape its development so that it's safe and beneficial for society.
For the polished, corporate version of me: see my resume
To see what I have worked on: see my projects
For everything else I'm proud of: see my activity
Otherwise here's the short version. I love:
Solving (hard) problems
chasing simple, elegant solutions
Creating stuff
working on my latest painting
Teaching and public speaking
lighting up when I talk of AI safety
Collaborating on meaningful projects
hackathon team designing AI institutions
Learning new things
(that's me fixing a plane before I fly it)
Pushing my limits
competition? challenge? hell yeah!
Featured
A glimpse of my recent projects and activities
CoT Monitoring Can Be Unreliable in Implicit-Influence Settings
CoT monitors often catch models explicitly instructed to hide their reasoning, but detection drops sharply when the same behavior shift comes from implicit cues instead. Preprint, under review.
OS-Harm: A Benchmark for Measuring Safety of Computer Use Agents
Presented at the ICML 2025 Workshop on Computer Use Agents. Spotlight at NeurIPS 2025 (Dataset and Benchmark Tracks)
Let's get in touch!
Looking to collaborate on a project? Need feedback or want to discuss ideas from my field? Exploring new opportunities or potential roles?
I'd be happy to collaborate, share insights, and exchange ideas: don't hesitate to reach out!