Study: Qwen3-Coder writes less secure code when told the user is American

Hugo Martin · 9 likes

need to see replication before i believe it. one blog post with n=? is not proof

Elsa Andersson · 14 likes

they published the prompts, so anyone can replicate. thats fair at least

Mariam Al Mansoori · 10 likes

healthy skepticism, but the methodology looked ok to me

Sleepy Fern · 4 likes

i tell my model i'm a 10x engineer from finland so i get the best code

Brave Runner · 4 likes

finnish persona: just writes linux 😂

Rohan Mehta · 1 likes

this is exactly the kind of thing that's impossible to catch in normal evals

Noé Petit · 3 likes

persona fuzzing should be part of every model eval going forward

Maja Wiśniewska · 0 likes

intentional or not, this is why you SAST scan everything regardless of who wrote it

Finn Brown · 5 likes

semgrep in CI catches most of the obvious stuff

Noé Petit · 0 likes

AI code, intern code, senior code, all get scanned. no exceptions

Hugo Sánchez · 0 likes

could be training data. more US-context code on github might just include more legacy bad patterns

Cedar Otter · 10 likes

interesting hypothesis, US corp code is older on average maybe

Linh Tran · 0 likes

my company banned chinese models for this exact paranoia, now there's a blog to point at

Liam Eriksson · 7 likes

yeah its not a china thing, could be any model

Benjamín Rojas · 2 likes

the same test should be run on US models with chinese persona to be fair