Paradigm Shock

The episode surfaces a few genuinely valuable ideas - LLM-generated code introducing more vulnerabilities while developers over-trust it, and ML model deserialization as an underappreciated attack surface with real Hugging Face payloads found. However, these are buried under heavy biographical framing, a World Cup segment, and the endlessly recycled 'it's a race' metaphor that substitutes for deeper analysis.

“the group that used AI assistants to write code, they wrote more insecure code and they believed that their code was secure because they trusted the models”

“we studied the serialization of models...we found that majority of those models that are out there, at least, uh, as of the time of the study, they were using unsafe, uh, serialization formats”

The deserialization-of-ML-models angle and the paper using smaller specialized guard models to protect larger ones are genuinely non-obvious contributions, but the broader framing collapses into stock security discourse - 'it's a race,' 'defense in depth,' 'Swiss cheese model' - that offers little a knowledgeable B2B operator hasn't already encountered.

“the race is much faster. We went from you know, cards to formula One cars. That's essentially it”

“security is something that should be layered...Swiss cheese, right? If you slice a Swiss cheese, they have holes in it”

Dr. Santos is a legitimate researcher with published work, a Google Research Scholar award, and hands-on experimental results (including accidentally hacking herself demonstrating a real exploit), making her clearly more than a thought-leader. However, she is an early-career academic with limited industry deployment experience, and several claims are delivered at a speculative or hypothesis level rather than from operational authority.

“I ended up deleting all my files in my machine. I ended up hacking myself due to a mistake that I have made”

“we did publish a paper very recently, I think last year on exactly the preliminary results on it”

There are genuine specifics: named papers (Seneca), the Java deserialization context, the Hugging Face reverse-shell findings, and a cited developer study of roughly 47-50 participants split by AI assistant use. But quantification is consistently loose - 'majority of models,' unprecised patch-window timelines, and the Mythos/Anthropic framing contains factual imprecision - limiting how actionable the evidence is.

“a study that was I think back in 2022 or 2023 where they interviewed I think 47 or 50, uh, developers”

“majority of those models that are out there, at least, uh, as of the time of the study, they were using unsafe, uh, serialization formats”

The host occasionally constructs genuinely layered follow-ups - flipping the Mythos framing to the offensive risk, probing the psychological over-trust finding, and pressing on specialized vs. general-purpose models - but undermines the episode with extended biographical segments, World Cup filler, unchallenged vague claims, and overt flattery that softens what could have been a sharper technical dialogue.

“I'm going to flip this around from a risk standpoint...are should we be, how concerned should we be”

“You're a big deal now. Your papers, you probably”