The brief · 22 July 2026
A security issue surfaced while testing a model hosted on Hugging Face, ahead of OpenAI's public writeup. If you're running third-party models hosted outside the big closed platforms, it's worth auditing your own eval pipelines before trusting a displayed benchmark score.
practitioners › Disclosed July 21; discussion hit 833 points and 576 comments on Hacker News.
Axios →Sponsored responses will start showing up in your conversations, alongside or after the generated content. For daily ChatGPT users, this shifts the product's nature: free-but-ad-supported rather than pure subscription.
practitioners › Program announced at ads.openai.com, 461 points and 375 comments on Hacker News, rollout gradual with no precise timeline yet.
ads.openai.com →The program bundles training and automation into a single subscription built for entrepreneurs without a technical team. If you're running your business solo and using ChatGPT in a casual, ad-hoc way, this is a chance to actually structure your usage instead of piecing together prompts as you go.
practitioners › Built on ChatGPT Work, targeted at small and mid-sized businesses, rollout announced with no confirmed availability date for France yet.
OpenAI →Des chercheurs ont construit un cadre statistique qui simule différentes réponses à une épidémie (intervention précoce, tardive, vaccination) et calcule leur effet sur le nombre total d'infections et la durée avant extinction. Ce n'est pas un outil grand public, mais ce genre de modèle nourrit les décisions de santé publique que tu subis ensuite, comme les seuils qui déclenchent un confinement ou une campagne de vaccination.
arXiv (q-bio.PE) →