AI news — Monday, 7 September 2026

AI’s day-to-day momentum is showing up less in flashy demos and more in the boring machinery: models you can ship, permissions you can trust, and software that plugs into real systems. IFM (MBZUAI’s Institute of Foundation Models) released K2 Horizon, a set of six Apache 2.0 open models ranging from 0.9B to 375B-A23B, trained on roughly 20 trillion tokens. The release also comes with a rare kind of accountability artifact: IFM says an audit of its own benchmark runs removed 24 trials across 10 tasks, nudging reported accuracy from 70.2% down to 66.9%. In other words, the paper trail is part of the product now.

That practical bent shows up elsewhere. A permissions-focused argument in retrieval-augmented systems warns that access control shouldn’t be bolted on after documents are fetched; if identity and authorization aren’t enforced during context assembly, a user who leaves a team can still have their “recent” queries pull finance documents for up to 17 hours, and later summarization may ignore updates. Meanwhile, startups are funding implementations: Jaipur Robotics raised €4.3M for 99%-accurate hazardous-material detection in waste-to-energy plants, and Cato secured €6M to monitor 27,000+ public tender sources in real time. Even Bybit’s new trading API—letting chat-based assistants execute orders via REST calls signed with user-held keys—pushes AI toward systems with tighter, more explicit control surfaces.