Fwd: [security-lunch] Jacob Lagerros on "Preventing AI model-weight exfiltration with FPGAs and inference verification — with live demo!"

[Alan Karp]

--------------
Alan Karp

---------- Forwarded message ---------
From: Michael Leo Paper <mpa...@stanford.edu>
Date: Mon, Sep 29, 2025 at 11:18 AM
Subject: [security-lunch] Jacob Lagerros on "Preventing AI model-weight exfiltration with FPGAs and inference verification — with live demo!"
To: securit...@lists.stanford.edu <securit...@lists.stanford.edu>

Security Lunch 🍂 Ed. — Wednesday,  Oct 1st, 2025, 12:00 pm @ CoDa E160

Preventing AI model-weight exfiltration with FPGAs and inference verification — with live demo!

Jacob Lagerros

Can't make it in person? Join us on zoom.

See our past & upcoming events on our website! 

Abstract: 

AI intellectual property is on track to become some of the most valuable IP on earth. To steal and reproduce a frontier AI model, an attacker would need access to its weights, its code, and a dozen accelerators. Code and compute is easy to get — but the weights can be TB in size, live in the data center, and have no reason to ever leave their storage / accelerators. The defenders options are much better. However, a frontier clusters needs to serve inference traffic to tens of millions of users. How can one simultaneously lock down AI model weights, while allowing such a massive volume of inference? This talk will introduce a novel solution, combining FPGA middle-boxes and machine-learning techniques for verifying the integrity of inference traffic. We aim for this solution be at once: 1) highly performant, 2) highly secure against very sophisticated adversaries, and 3) highly retrofittable to modern clusters with various bespoke architectures and optimizations.

Bio:

Jacob Lagerros is the CEO of Ulyssean, an AI startup building security and verification for large AGI clusters, backed by leaders from Anthropic, DeepMind, Meta, CrowdStrike, and more.

_______________________________________________
security-lunch mailing list
securit...@lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/security-lunch