EPFL AI Center x UK AI Security Institute - mini-symposium — Evaluations, Impact, and the Science of AI Safety

Join us at the EPFL AI Center on Monday, September 29, for a mini-symposium—“Evaluations, Impact, and the Science of AI Safety.”

Organized with the UK AI Security Institute (AISI) and the EPFL NLP Lab, the program will dive into AISI’s current priorities, methods for sandboxing agentic evaluations, fresh evidence on the real-world impacts of AI advice, and the science of evals. We’ll wrap up with a poster session and networking apéro.

Bring a poster—the apero session will highlight EPFL student work. When registering, select “Bring a poster” and add your title (limited slots).

! Registration is required to attend this event. Access is restricted to EPFL participants. !
Registration form (restricted to EPFL email addresses): https://forms.office.com/e/wPBhzPbzFP

Program

• 14:30 – 15:00 — Welcome and check-in
   
• 15:00 – 15:05 — Welcome, Prof. Antoine Bosselut, EPFL

• 15:05 – 15:20 — Opening & overview of AISI, Shannon Yang, UK AI Safety Institute (DSIT)
   
• 15:20 – 15:40 — Sandboxing agentic evaluations, Art O’Cathain, AISI
   
• 15:40 – 16:00 — “AI use in political information search”, by Lennart Luettgau
   
• 16:00 – 16:15 — Short break
   
• 16:15 – 16:35 — The science of evaluations, Magda Dubois, AISI
   
• 16:35 – 17:00 — "Fundamental Limitations in Defending LLM Finetuning APIs" by Eric Winsor

• Agentic evaluations & sandboxing frameworks
• Measuring real-world impacts of AI advice & wellbeing
• Science of evals: robust, transparent benchmarking
• Trustworthy AI practices: reproducibility, transparency, ethics