What does truly secure inference look like for superintelligent systems? This project prototypes a minimal, SL5-aligned inference runtime plus one or two unconventional, no-network I/O paths (serial, optical, audio FSK, maybe even printing) as a micro cross-domain gateway for air-gapped inference. We will threat model and evaluate the design against SL5 risks like model weight theft, sabotage, and covert exfiltration, and quantify the attack surface of μInference versus a standard ML stack.
About the project
Frontier AI labs will need ways to run and interact with their most sensitive models under extreme threat assumptions. Today’s inference stacks are huge, fragile, and full of hidden attack surfaces. In this project, we aim to remove most of the stack instead of piling defenses on top of it.
You will help build and test a minimal inference runtime and a secure I/O gateway that avoids IP networking entirely. The core work is to prototype one or two unconventional I/O paths, integrate them into a tight inference loop, and measure what we gain in auditability, code size, and exfil resistance. Then we will map findings directly to SL5 threat models and produce evaluation metrics that make the security tradeoffs legible to real operators.
If you are excited by weird-but-practical security engineering, this is a rare chance to build something concrete at the boundary between AI safety and high-assurance systems.
Theory of change
This project aims to deliver a concrete candidate for inference that a frontier lab could plausibly deploy for their most sensitive models, like how would you talk to a misaligned superintelligence in a secure way.
More broadly, it instantiates SL5’s radical minimization idea: instead of adding more layers on top of fragile stacks, we remove most of the stack. The outputs will include a prototype micro-stack and a reusable framework for reasoning about code size, auditability, and exfil paths that can plug into SL5 network and software-hardening specifications.
Your role
Mentees will study and reproduce the μInference proof of concept, then propose and execute at least one additional minimization or I/O experiment. You will implement and benchmark prototypes across one of three tracks (A/B/C) and compare results against a standard inference stack. You will also write short design docs and threat models that tie your results back to SL5 standards across network, machine, supply chain, and personnel security.
We will collaborate through a weekly 1-hour sync, a weekly 1-hour retro, and async chat. The minimum commitment is 5 hours per week, but the expected commitment for strong progress is closer to 10 to 15 hours per week. I will take up to 3 mentees and provide detailed guidance, including 2+ hours per week per mentee.
Prerequisites
You should have a security mindset, be comfortable with open-ended exploration, and be willing to document findings clearly. You should have ML basics, meaning you can run simple inference loops and reason about model inputs and outputs. Experience with systems programming (C/Rust) or OS and infrastructure work is helpful but not required. The most important trait is curiosity and seriousness about AI security, SL5, and building weird but practical prototypes that could matter to real labs.
Application question(s)
- Our PoC compiles to ~1M LOC vs ~27M LOC when we clone the full tree. Name three specific subsystems or layers you would remove or collapse next for SL5 inference, and explain why.
- Propose one unconventional, SL5‑compatible way to get prompts/responses in and out of an air‑gapped μInference box without IP networking. Why is it promising? What are the biggest risks at SL5?
About the mentor
Luis Cosio is a Mexico-City based technologist and entrepreneur with 15 years at the intersection of cloud, cybersecurity, and artificial intelligence. He has architected and launched national-scale systems including Mexico’s largest e-government project that serve millions of users every day. After founding, scaling, and exiting multiple tech companies, he is now pursuing an M.S. in AI at Johns Hopkins University while working full-time on AI-safety and alignment research, with an emphasis on SL5.
Drawing on experience that spans from national infrastructure projects to working with frontier AI labs, Luis mentors with a focus on high-impact technical projects. Mentees can expect hands-on guidance in transforming research ideas into deployable prototypes, integrating technical rigor with practical considerations.