What we are doing, and why

What We Do

Intelligence Security Laboratories is a nonprofit research and development lab working on the security of advanced AI infrastructure. We design, build, and test high-assurance systems that protect critical AI models and the infrastructure that runs them.

ISL is designing and prototyping a secure inference data center: a facility that keeps the frontier model inside it beyond the reach of any entity that would steal, alter, or misuse the model. That includes adversaries with the resources of a nation-state, other highly capable AI systems, and the model itself. To our knowledge, no facility in the world meets that standard today. For more details on the design and methodology, see the RAND report Highly Secure Inference Data Centers ↗, co-authored by ISL’s executive director, Gabriel Kulp.

Why It Matters

On current trends, a transformatively capable AI model is only a handful of years away, and with it, escalating risks. If a frontier model can be stolen or altered, the safeguards built into it can be removed. Further, a frontier model able to plan and act may become an adversary in its own right.

The work to ensure these models are secure must be finished before they arrive. We expect this to be difficult in ways that are hard to anticipate. By confronting these challenges early, we build practical knowledge and uncover problems with enough lead time to solve them.

The standard approach to AI security is to take the best available products and harden them further. Commercial security assumes breaches will happen and then focuses on detecting and containing them afterward. For frontier models in the near future, the impact of such breaches could be severe, and the most dangerous attacks are the ones no defender has thought of. A single undetected compromise could be irreversible, so the security must be inviolable.

A security protocol with this degree of efficacy has not yet been demonstrated. We assert that withstanding unforeseen attacks requires designing a facility around provably secure conditions rather than anticipating and addressing weaknesses. We would much rather have the first serious attempt with this strategy happen long before it is strictly required; ISL is building this rehearsal.

How We Work

Designing against unknown attacks. ISL uses Systems-Theoretic Process Analysis for Security (STPA-Sec), a systems engineering methodology that shifts the question from which attacks to anticipate to which conditions the system must never permit. We apply formal methods in standard and novel ways to analyze these conditions and constraints. Our standard is no-fail rather than best effort, which means that our work diverges from usual enterprise software and data center security work and instead looks more like avionics, reactor control, and root-key ceremonies ↗.

Owning the whole system. A facility can have secure hardware, strong cryptography, insider mitigations, and strict physical access controls, and still have weaknesses in how those parts work together. ISL solves this problem by designing, analyzing, and prototyping the whole gold-standard facility as one system.

Building, not only writing. Previous work in this area was predominantly theoretical. It has done an excellent job describing the problem and outlining solutions. However, designs that have never been assembled hide unexpected weaknesses. ISL is building at a scale large enough to uncover design oversights and practical challenges while there is still time to fix them.

Outputs and Impact

We are developing a wealth of new knowledge and expertise in a space that is increasingly important. Our work demonstrates what is technically feasible and gives decisionmakers concrete results and proposals to evaluate.

ISL’s goal is not building or operating the world’s high security data centers. Rather, we produce the knowledge and technical artifacts to share with transition partners.

Philanthropic funding allows ISL to remain independent of the incentives of a particular customer or operator and to instead purely pursue work in the public interest.