AI Red Teaming Isn't Enough: Shoshana Cox, AI Security Researcher

For years, security teams have relied on testing to validate defenses. With AI systems, that logic breaks.

The reason is mathematical. Every AI model carries what researchers call an adversarial subspace: a set of inputs that will cause it to fail. These spaces are massive, sparsely clustered, and computationally infeasible to search. You cannot enumerate all the attacks. Which means you cannot test for all the attacks. Which means spraying thousands of prompts at a model and blocking the ones that work is not security. It is a demonstration.

Shoshana Cox has been working on AI security since 2010, before generative AI existed. She wrote the first machine learning security operations paper in 2022. She holds a patent in federated AI architecture. She served on the core author team of the OWASP AI Exchange and helped prepare the technical requirements for the EU AI Act. She has spent years watching the industry build products that sell confidence without delivering protection, and she has paid a professional price for saying so publicly.

In this conversation, Shoshana explains what adversarial subspaces actually are and how they form. She walks through attack transferability: how an attacker does not need access to your model to build attacks that work against it. She breaks down why AI monitoring with defined thresholds is how you deploy AI securely. And she makes the case that threat modeling is not optional groundwork. It is the only starting point that makes the rest of the work meaningful.

This is a technical conversation. It does not offer quick fixes. What it offers is a clearer picture of how AI systems actually fail, and what it takes to defend them seriously.

For security architects, CISOs, and anyone responsible for AI deployment decisions.

Guest: Shoshana Cox, AI Security Researcher, OWASP AI Exchange Core Author, and former Red Teamer.