Strategic Context

On current trends, we expect a transformatively capable AI model in a handful of years. The developer of that model should secure their infrastructure: an insecure model could be tampered with or stolen by adversary nation-state actors, other AIs, or itself. Navigating transformative AI safely may depend on frontier models remaining secure from tampering and theft.

We expect that, in making their infrastructure secure, they will reach for the most available off-the-shelf solutions and that these solutions will likely fail. Highly capable adversaries will almost certainly find and exploit flaws, and we cannot count on noticing this happening.

What is ISL doing?

ISL exists to develop and demonstrate the required level of security for critical AI deployments in advance of when it’s needed. We expect this to be difficult in ways that are hard to anticipate. By confronting these challenges early, we uncover problems with enough lead time to solve them and build key practical knowledge before it’s required.

We think that this is hard and requires a new focused research and development effort. Even if each component is investigated deeply, the secure data center, as a holistic system, will not have been proven secure. Someone must own the whole problem as an integrator, and we aim to fill that role.

We think that solving this problem requires methodological innovation. We are applying STPA-Sec in a way it wasn’t quite designed for. We are leveraging AI design tools without compromising the integrity of the design process. We are using formal methods and other high-assurance approaches everywhere we can. This project is more like avionics or reactor control than enterprise software or red teaming.

STPA-Sec (Systems-Theoretic Process Analysis for Security) makes this feasible. It is a system design methodology that provides an emphasis on unknown-unknown attack vectors. See the RAND report Secure Inference Data Centers: A Vertically Integrated Strategy for Security Engineering for details.

If this sounds exciting to you, we’re hiring!