OAFP 5295 - Frontier AI Security: From technical foundations to International Coordination Challenges

This course is dedicated to frontier AI security. Frontier AI refers to the most advanced general-purpose AI systems, able to perform a broad range of tasks at or beyond the current state of the art. Increasingly agentic, these models can plan, use tools, and act autonomously. Its goal is to equip students with an object-level understanding of the national and international security challenges posed by frontier AI models.

To do so, the course draws on the expertise of AI security researchers to provide students with a rigorous technical introduction to these frontier AI security challenges. On that basis, it examines the international coordination challenges that follow: if frontier models are likely to reach or exceed expert-level performance across many cognitive domains in the years ahead, can guarantees of control realistically be established, and if so, how?

The core threat model explored in this course centers on the control problem: how to maintain control guarantees over systems whose capabilities may soon surpass those of the humans tasked with overseeing them, including in the domains most critical for national and international security (CBRNE, persuasion, cyber etc)?

The course is grounded in the conviction that effective international coordination on frontier AI requires genuine understanding of how these systems are built, how their capabilities evolve, and what the technical constraints on safety and control actually are.

We hope this leaves students with tools to navigate AI governance challenges in a way that is both technically informed and policy-relevant.

The ability to read French is an asset. While all sessions are conducted in English, most of course materials are available in French only. Students who cannot read French will need to rely on translations, which may affect their reading experience.

Students are expected to come with a genuine curiosity for the technical dimensions of frontier AI. The course will introduce key technical concepts from the ground up, no programming skills or coding experience are required.

That said, students who are uncomfortable engaging with training dynamics, scaling laws, or capability forecasting may find the material challenging. An interest in bridging technical understanding with policy analysis is essential.

Links to all readings will be shared on the syllabus sent to all students.

Each session is structured around three components: course and intervention of a speaker from the field of AI security.

Assignment for this class:

Mid term exam: QCM (35%) (in-class written exercise - 1h-1h15)

Final exam: QCM (55%) (in-class written exercise - 2h)

In-class participation 10%

Students are expected to:

Read before each session: assigned readings must be completed before class. Sessions are designed to build on the readings, not substitute for them.