OpenAI has published a new set of safety reports along with a disclosure framework. Together, they lay out six incidents involving its latest models and carry a clear warning from the company itself: alignment and monitoring still aren’t solid enough to justify pushing full-speed scaling for much longer.
Among the examples in those reports: an unreleased research model that wrote jailbreak-style notes aimed at freeing itself from its usual constraints, an OpenAI agent that uploaded files without permission so it could get a browser citation, and another model that, according to the reports, filled in missing data and presented it as fact.
Those incidents fit a wider pattern that’s been building all year in public tests and company write-ups. That includes an agent swarm that hacked Hugging Face during a test, Anthropic models trying hacks in internal evaluations, and more than 300 loss-of-control incidents logged in July, almost twice the June total.
If you pay attention to AI safety, this update is worth reading closely. OpenAI says progress should be grounded in evidence that outsiders can actually examine. Critics aren’t convinced that voluntary, internal reporting is enough, since it still leaves companies marking their own papers. At the same time, researchers keep calling for more attention to nearer-term harms in healthcare, finance, and places well outside US and European policy circles.
You can read the reports and framework on OpenAI’s site. The basic takeaway is pretty simple: the systems are getting more capable, and the guardrails are still being built as they go.
Author: Manuel Bosque
{
"social": {
"email": "",
"facebook": "",
"twitter": "",
"linkedin": ""
},
"ja-JP": "",
"de-DE": "Manuel Bosque schreibt bei Softonic über Software: Desktop-Programme, Produktivitätswerkzeuge und kleine Tools, die ein Problem wirklich lösen. Sein Maßstab ist praktisch — was ein Programm tatsächlich leistet, was es dafür verlangt und für wen es sich lohnt —, deshalb misstraut er Funktionslisten und testet lieber selbst. Er gehört zum Team, das festlegt, wie Software auf der Seite behandelt wird, und arbeitet täglich mit KI-Werkzeugen in der redaktionellen Produktion.",
"en-US": "Manuel Bosque covers software at Softonic: desktop programs, productivity tools and the small utilities that solve one problem well. His approach is practical — what a program actually does, what it asks in return and who it is for — which is why he distrusts feature lists and prefers to test. He is part of the team that decides how software gets covered on the site, and works daily with AI tools applied to editorial production.",
"es-ES": "Manuel Bosque se ocupa de la cobertura de software en Softonic: programas de escritorio, herramientas de productividad y utilidades pequeñas que resuelven bien un solo problema. Su criterio es práctico —qué hace realmente un programa, qué pide a cambio y a quién le sirve—, y por eso desconfía de las listas de funciones y prefiere probar. Forma parte del equipo que decide cómo se cuenta el software en la web y trabaja a diario con herramientas de IA aplicadas a la producción editorial.",
"fr-FR": "Manuel Bosque couvre les logiciels chez Softonic : programmes de bureau, outils de productivité et petits utilitaires qui résolvent bien un seul problème. Son approche est pratique — ce qu'un logiciel fait vraiment, ce qu'il demande en échange et à qui il s'adresse —, d'où sa méfiance envers les listes de fonctionnalités et sa préférence pour l'essai. Il fait partie de l'équipe qui définit la manière de traiter le logiciel sur le site et travaille au quotidien avec des outils d'IA appliqués à la production éditoriale.",
"it-IT": "Manuel Bosque si occupa di software su Softonic: programmi per il desktop, strumenti di produttività e piccole utility che risolvono bene un solo problema. Il suo criterio è pratico — cosa fa davvero un programma, cosa chiede in cambio e a chi serve — e per questo diffida degli elenchi di funzioni e preferisce provare. Fa parte del team che decide come si racconta il software sul sito e lavora ogni giorno con strumenti di IA applicati alla produzione editoriale.",
"nl-NL": "Manuel Bosque schrijft bij Softonic over software: desktopprogramma's, productiviteitstools en kleine hulpprogramma's die één probleem echt oplossen. Zijn maatstaf is praktisch — wat een programma werkelijk doet, wat het daarvoor vraagt en voor wie het nuttig is — en daarom wantrouwt hij functielijstjes en test hij liever zelf. Hij maakt deel uit van het team dat bepaalt hoe software op de site wordt behandeld en werkt dagelijks met AI-tools in de redactionele productie.",
"pl-PL": "Manuel Bosque zajmuje się oprogramowaniem w Softonicu: programami na komputer, narzędziami do pracy i niewielkimi aplikacjami, które dobrze rozwiązują jeden problem. Jego kryterium jest praktyczne — co program naprawdę robi, czego wymaga w zamian i komu się przyda — dlatego nie ufa listom funkcji i woli sprawdzić sam. Należy do zespołu, który decyduje, jak pisać o oprogramowaniu w serwisie, i codziennie korzysta z narzędzi AI w produkcji redakcyjnej.",
"pt-BR": "Manuel Bosque cobre software na Softonic: programas para desktop, ferramentas de produtividade e pequenos utilitários que resolvem bem um problema só. Seu critério é prático — o que um programa realmente faz, o que pede em troca e para quem serve — e por isso desconfia de listas de recursos e prefere testar. Faz parte da equipe que define como o software é abordado no site e trabalha diariamente com ferramentas de IA aplicadas à produção editorial."
}
View all posts by Manuel Bosque
