Guy

Projects

Research Direction: Security Level 5 for AI

Building systems that can be outright secure against the most capable threat actors in the world is impossible. We tackle the merely very difficult problem of building such systems for the most sensitive parts of a frontier AI datacenter.

I'll be happy to mentor projects that promote this, such as:

• Software / hardware infrastructure to securely incorporate LLM agents into AI R&D processes

• Extremely simple (as opposed to complex, not easy) alternatives to existing complex solutions (e.g. FIDO2 key with ~10K transistors, code review system with minimal code)

• Building restrictive interfaces for ML libraries and protocols (e.g. limiting which instructions can be loaded in CUDA kernels)

• <Your idea here!>

See sl5.org for more information.

What I'm looking for in a Mentee

The best candidate would be someone well-versed in a relevant subject who wants to apply their skills to enable the security of future frontier AI.

You should be able to independently own a technical project, and you'll be given the information, access and resources you need for it.

What I’m like as a Mentor

Expect to mostly be working independently, with weekly meetings to make sure the project is on the right track and more-than-weekly minor reviews of intermediary outputs.

In addition, I'm available for frequent low-bandwidth communication.

Bio

Before researching SL5, I worked in several organisations as a HPC chip engineer, vulnerability researcher and R&D project lead. You can also check out my musical about AI, Out of This Box: The Last Musical (Written by Humans).