Intelligence Security Laboratories

Intelligence Security Laboratories (ISL) is a 501(c)(3) nonprofit research lab that describes its focus as "implementation-driven R&D for high-security AI systems" ¹. It states that it exists to develop and demonstrate the level of security that critical AI deployments need, before it is needed ¹.

ISL expects the developer of a transformatively capable model to reach for off-the-shelf security solutions that will likely fail ¹. It warns that an insecure model could be tampered with or stolen by nation-state actors, other AIs or the model itself ¹; see Model weights or data have not left the facility.

Its stated approach has five parts:

• ISL argues that studying each component in depth will not prove a secure data centre secure as a whole system ¹. It aims to be the integrator that owns the whole problem ¹.
• It applies STPA-Sec (Systems-Theoretic Process Analysis for Security), a system design method that emphasises unknown-unknown attack vectors ¹.
• It uses formal methods and other high-assurance approaches wherever it can ¹.
• It expects demonstration infrastructure to help train relevant talent, show policymakers what is feasible, and support discussion of high-assurance frontier AI ¹.
• It plans to hand the knowledge and technical artifacts it produces to other actors when needed, and states that it does not expect to become the world's producer of secure compute ¹.

For details of STPA-Sec, ISL points to RAND's secure inference data center design ¹. Its executive director, Gabriel Kulp, is a co-author of that report ^{3 2}. Before founding ISL, he was a fellow at RAND working on hardware-enabled governance mechanisms for GPU export controls and on international verification of agreements ². Tom Gardiner is ISL's staff director and Paul Murley its technical director ².