Time: November 17, 2026  |  1:00 pm ET  |  10:00 am PT
Speakers:
Brian Richardson, Director of Services
Chase Snyder, Director of Product Marketing

Eclypsium researchers discovered a way to change the output of a running AI model without touching the model file, the prompt, or the application. This type of model attack bypasses conventional integrity checks by compromising the underlying infrastructure, making this malicious attack look like a hallucination.

In this session, Eclypsium's Brian Richardson walks through the new research:

• How the attack works, why existing integrity checks miss it, and why a manipulated answer can look exactly like a hallucination. He closes with the configuration checks and questions for AI hardware providers that operators can act on now

You'll learn:

• Which components inside an AI server can influence a model while it runs
• Why hashes, signatures, and attestation don't catch runtime manipulation
• What to check on your own AI servers today

Unable to attend?  Go ahead and register and we will send you the links to the recording and slides.