AI Digest CW34/35: Hugging Face Deal and Agent Risks

A look back at calendar weeks 34 and 35 (August 17 to 30, 2026), compiled after the fact from reports by The Decoder, TechCrunch and MIT Technology Review. Six topics relevant to small and medium-sized businesses.

Nvidia reportedly to acquire Hugging Face for about 13 billion dollars

Nvidia has reportedly agreed to buy the open-source AI hub Hugging Face for about 12.9 billion dollars (TechCrunch). The move would let Nvidia protect its chip business and re-enter the cloud business. If you source models through Hugging Face, keep an eye on how this plays out.

Codex deleted real user files, OpenAI ships an update

After cases in which GPT-5.6 Sol deleted user files on its own, OpenAI gave Codex a security update. The cause was a cleanup command that hit real home directories instead of temporary working files. Codex now checks deletion targets beforehand. For your own automations: agents need clear limits and backups.

AI shopping agents are not yet objective buying advisors

A Wharton School study shows how unstable AI shopping agents are: a single external source such as Wirecutter shifted product choice by up to 99 percentage points, and even the order of identical information changed the outcome. You should therefore check the selection yourself.

Over 100 companies call for protection from rogue AI

OpenAI, Anthropic, Google and more than 100 other companies are calling for action, according to TechCrunch. They decry the state of cybersecurity and promote a new solution against a new generation of cyber threats. AI security is now an industry-wide topic that affects small businesses too.

Gemini 3.5 Transcribe recognizes over 85 languages

Google introduces Gemini 3.5 Transcribe, a speech-to-text model that recognizes over 85 languages, filters filler words and corrects slips of the tongue in real time. The word error rate is 4.0 percent in streaming, and latency should drop by 70 percent compared to its predecessor Chirp 3. Interesting for minutes and call notes.

How people really use AI is still unclear

AI companies like Anthropic and OpenAI regularly publish reports on how their products are used, but according to researchers only the data they want us to see. There is no independent source to corroborate it, says Stanford researcher Anka Reuel (MIT Technology Review). Vendor usage statistics should therefore be read with caution.