CloudNext

The core argument - that adversarial subspaces are computationally intractable to enumerate, making prompt-spray red teaming security theatre - is genuinely non-obvious and well-articulated. The surrogate model attack transferability discussion adds real depth. However, roughly half the runtime is consumed by tangential banter (karma, AG1, McKinsey, women in tech) that contains no actionable insight for a B2B operator.

“these boxes, these spaces, are absolutely massive. We've determined that it's not feasible computationally to run a search through them...what you can do, because there are effectively an infinite number of attacks, is spray a bunch of prompts at something and say, oh, my God, see, look, all these attacks worked. And then, look, I blocked them all. Now you can do that. But did you actually really provide meaningful security? Uh, I would argue no, you absolutely did not.”

“attackers in the wild don't need your model, they don't need your data, they don't need any of that. They need to understand the domain that your model is trained on in order to construct these attacks”

The reframing of AI red teaming as a mathematically broken practice - not merely imperfect - is a genuinely contrarian and well-grounded position. The adversarial subspace geometry explanation and centered kernel alignment application are underexplored publicly. The standard advice (threat model first, don't deploy AI if you don't need it) is less original and pulls the score down.

“Every model comes with, um, a set of attacks that will work against it...it's called a subspace. An adversarial subspace...You can think of it like a box that's attached to the model”

“they get there because whenever you compress a high dimensional space into a lower dimensional one, um, you break certain assumptions, right? You create angular representations among, uh, the concepts that the model so quote, unquote learns”

Cox has verifiable, hard-to-fake credentials: a 2022 first-of-kind ML security ops paper, a 2024 federated defence patent, core authorship on OWASP AI Exchange, and contribution to EU AI Act technical sections. She is clearly a practitioner who has done the work, not a recycled thought leader. Score is not higher because she is early-stage (stealth startup) and much of her most specific work is undisclosed.

“I wrote a paper in 2022 that was the first, uh, AI security operations paper in machine learning, security operations...with the first architecture for exactly how to implement this defensive, um, security. And then I got a patent, um, in 2024 for defensive, uh, federated architecture”

“I'm on the team that sets red teaming requirements”

There are some concrete data points - named technique (centered kernel alignment), named collaborator, course price ranges ($1,500 - $3,000/head), and the 2022/2024 publication dates. But the guest is deliberately vague about company names, paper titles, and threat actors for legal reasons, and the key threshold example is explicitly fabricated ('this is a meaningless number'). No named vulnerable systems, no published attack success rates, no cited studies.

“I know for a fact that these guys selling courses out here are like 2000, 1500 to $3000 ahead, basically. And you pack a course with like 20 people. You do this multiple times a year. You are raking in cash”

“Let's say I have a ratio of red blots, red dots to blue dots going through my system...Let's say I have a ratio...becomes um, 27% higher for the red ones...I will make up a magic number. This is a meaningless number.”

The host repeatedly derails technically rich moments into comedy tangents (karma, Tower of Babel, AG1, gold), openly admits he cannot follow the technical content ('I'm too thick for that information to go in'), and never pushes back on or probes any specific claim. The women-in-tech and personal anecdote sections, while humanising, are handled without any craft or follow-up rigour. A few functional questions ('Is it ever solvable?', 'What does that mean in real world terms?') prevent a lower score.

“I can't even imagine how that will apply in the real world. I'm too thick for that information to go in, but beautifully presented.”

“Um, definitely interested. Should I buy silver? Gold.”