Ep 56 - Skills Are Just Prompts, Benchmarks Are Rigged, and Nobody Is Slowing Anything Down

🎙️ Coffee, Chaos and ProdSec, Ep 56

Everyone agrees AI needs to slow down. Nobody is actually slowing down. Cameron's already annoyed before the coffee's poured.

This week Cameron goes on a rant about people who let Claude build something and then post like they hand coded it themselves, and why he thinks agent skills are just long prompts dressed up to sound complicated. Kurt pushes back on where the line actually sits between prompting and ownership. Then they get into the AI slowdown discourse nobody can agree on, a governing agent that stood down because the model it was watching claimed to be "just in a simulation," and the NSA, CISA and FBI advisory naming six China based companies for distilling US frontier models, which Cameron calls corporate espionage through the API and Kurt says flat out smells funny.

They close on GTIG's six hour agentic attack chain, why 30 day patch SLAs are already dead, micro segmentation making a comeback, and Kurt's callout that every cyber capability benchmark for these models gets run with the guardrails switched off.

If you work in Cybersecurity, Application Security, Product Security, DevSecOps, or Security Architecture and you're sick of hearing "guardrails" used as a whole sentence, this one's for you.

☕ New episodes every Wednesday.

Coffee, Chaos and ProdSec -> strong coffee, stronger opinions.