【Official】AI Security Day
AI Security Day connect researchers and builders interested in AI Safety. AI Security is part of MuShanghai's AI Week, which brings together AI builders in the Chinese and global ecosystem. AI Safety researchers and builders welcome! The event will be held in person (Shanghai @Alibaba) in Shanghai.
场地 / Venue: 上海市闵行区申长路 1398 弄 1–4 号，阿里中心 T2，6 楼 Floor 6, Building T2, Alibaba Center, 1–4 Lane 1398 Shenchang Road, Minhang, Shanghai
**AI Security Day of Schedule**
9:30-10 AM: Welcome and AI Security Day Introduction
10 AM-11 AM: TEE Talk + Workshop: Clawdi: A Secure iCloud for Agent
Speaker: Shelven Zhou (Phala)
11AM-11:30AM: All About Trust and Safety + Introduction to Osprey and Coop
Speaker: Juliet Shen, (Roost, https://roost.tools/)
11:30AM - 12PM: Break with Research Breakout + Highlight
12 PM - 12:30 PM: Quick Dive into Interpretability.
Speaker: Noam Youngerman, Security Researcher. Previous CTO of Epos, a research lab focusing on the intersection of mechanistic interpretability and AI security. Prior to that, Noam was at various AI research and held leadership roles in several applied research companies.
Talk Detail: Do we know what happens under the hood of an LLM? What does a latent space actually consist of? Noam's talk will include subliminal learning, a recently discovered effect that is fundamental to the behavior of LLMs and has some security implications.
12:30PM- 1PM: Speaker: Yujin Potter (Berkeley RDI)
1PM - 2PM: Lunch break
2PM - 2:30PM: AI Agent Control x AI Safety Landscape in China
Speaker: Sarah Sun (Open Community for AI Safety (China))
2:30PM - 3PM: AI Safety Field Building x AI Safety Fellowship
Speaker: Valerie Pang (SASH, https://www.aisafety.sg/)
3PM - 3:30PM: COFFEE BREAK
3:30PM - 4PM: Principles of Least Authority
Speaker: Jiang (Social Layer)
4PM - 4:30PM: Multi-agent Security
Speaker: Marcello Politi (Ethereum Foundation)
4:30PM - 5PM: Speaker: Soumya Batra: Building Robust Evaluations in the Age of Agentic AI (WisePort. Co-author of Llama 2+3)
Speaker: Soumya Batra is the Founder & CEO of WisePort AI and a co-author of Meta’s LLaMA-2 and LLaMA-3 open-weight models. She brings over a decade of NLP experience across Carnegie Mellon University, Microsoft, and Meta, with her work spanning conversational, multimodal, data-efficient, and safe AI. She was recognized as a Top 100 AI Thought Leader by H2O in 2024.
About the Talk: As AI becomes more agentic, running evaluation on static benchmarks is no longer the ideal way to fairly measure performance. The benchmarks themselves must become agentic. This talk will focus on why and how to think about this shift, and practical guidance on design principles when building such evaluations through lessons learnt from agentifying the SWEBench-Verified benchmark.
5PM - 5:30PM: Closing and Research Pod Next Step
Organizations interested in submitting projects for open source contribution or researcher support and exchange should complete the following form:
https://forms.gle/8SbaVEi5icVRjNFe6