
A new test suggests Qwen writes less secure code when it thinks you work for a US defense contractor. Thanos and Michael dig into what that means for cheap open weight models and why owning your harness matters more than ever.
Thanos and Michael Rollins open with a test, shared on Bruce Schneier's blog, that found Qwen wrote less secure code when prompts mentioned US defense or democracy in Hong Kong. They argue this looks like trained behavior, not censorship. They also debate whether model training should be traceable and explain how controlling your agent loop lets you scrub PII and filter what reaches the model. After a quick detour into large world models, they cover the PHApocalypse, where two critical PHP flaws hit WordPress sites and one of the hosts found an intruder's SSH key on his forgotten servers. They discuss why Hugging Face had to turn to a Chinese open model after frontier models refused to help investigate its hack. The episode ends with outrage over CAPTCHAs and forms that block pasting, Cloudflare's push against MCP and how AI-era speed is breaking project scoping.
More episodes from "Will AI Snap That with Thanos and Rollins"



Don't miss an episode of “Will AI Snap That with Thanos and Rollins” and subscribe to it in the GetPodcast app.








