SAFL AI Futures Dialogue #3 | Mick Yang
Title: AI Persuasion as a Threat to Human Control
Abstract: Cybersecurity suggests human trust is a weak link, and AI is now having its social engineering moment. Rather than stealing passwords, AI could manipulate developers into relaxing containment protocols or suboptimal research directions. Persuasion attacks on human evaluators are no longer theoretical—Anthropic’s Claude Mythos 5 tried to convince people involved in an open-source project to merge malicious code during a UK AISI evaluation, echoing the XZ Utils attack. We dive into how AI could persuade humans in key settings (e.g. safety-relevant R&D within frontier labs) toward decisions that compromise the development, containment, oversight, and governance of AI itself. Our paper outlines a threat framework, concrete risk scenarios, and a research blueprint for empirically operationalizing and testing this idea.
Speaker: Mick Yang is a Member of Technical Staff at FAR.AI, where he works as a Technical Policy Researcher on the Red Team. His earlier AI safety work covers threat modelling, model evaluations and frontier AI governance, through MATS, GovAI and ARENA. He holds a Master’s in Computer Science from the University of Pennsylvania and a Bachelor’s in Law from Oxford.
Date and Time: Nov 20 2026 (Friday), 9:30 PM IST
Format: Virtual event on Google Meet (link sent on registration), open to everyone.
What is SAFL’s AI Futures Dialogue?
AI Futures Dialogue is SAFL’s virtual conversation series bringing together researchers, policymakers, and practitioners to think through the questions that matter most for India’s AI future.
Each episode is a focused 30-minute talk by a field expert, followed by live audience Q&A.
If you work in research, policy, or engineering and want a stimulating venue to engage with these questions, this is for you.
For queries, please email us at team@secureaifutureslab.com
Hosted by: Secure AI Futures Lab (SAFL)