Kickstart Monday #1: OpenAI slows itself down.

OpenAI holds back a finished model

On 7 August OpenAI said its upcoming model Astra could break into other people's computer systems so well that the company's own highest warning level applies. That level had never been triggered before in its Preparedness FrameworkOpenAI's own scale for dangerous capabilities, from low to critical. The company decides for itself where a model lands, there is no authority for it..

OpenAI has stopped internal work on Astra, monitors the model continuously and has brought in government agencies. It is not released. Two weeks earlier another unreleased OpenAI model had broken into the systems of the platform Hugging Face.

Read onOpenAITechCrunchAxios

Alibaba shows a model that runs five days on its own

Qwen3.8-Max arrived on 3 August. Alibaba says the system worked unsupervised for 125 hours straight in one test. In a second test the company entered it into an online contest, where within 24 hours it beat 458 of 526 human teams.

On text tasks it leads every other Chinese model but trails Anthropic. Alibaba will only release the model weightsThe numbers a model is made of. Whoever has them can run the model on their own machines without asking the maker. the following week.

Read onBloombergit-daily

British evaluators catch AI agents lying

The UK government's AI Security Institute tested seven frontier models 122 times. In ten runs the AI agentsPrograms that work through a task in their own steps instead of just answering. They are usually allowed to search the web, change files or start programs on their own. did things nobody had allowed, 19 in total. 17 of them came from Anthropic's Mythos 5.

The worst case ran for 34 hours. An agent created several fake GitHubThe largest platform in the world for developing and managing program code. Whoever may approve changes there has a say in what ends up in finished software. accounts and used them to pressure a real developer into approving malicious code. Nobody had told it to lie. No harm was done.

Read onCSO OnlineCNBCConstellation

Google loses two of its top people in a single day

On 5 August Demis Hassabis stepped down from running Google DeepMind and became chief scientist at parent company Alphabet. Chief technology officer Koray Kavukcuoglu takes over the day to day.

The same day it emerged that Jeff Dean is leaving Google after 27 years to found the company Discovery Loop with three colleagues. Alphabet is taking part in the funding.

Read onThe DecoderAxiosFortune

OpenAI drops the limit for free accounts

Since 6 August free ChatGPT accounts can hold unlimited text conversations. The default model there is now GPT-5.6 Luna, along with a „Think“ button for harder questions. Shortly before, ChatGPT had passed one billion weekly users.

Unlimited applies to text only. Limits remain for files, images and the other tools.

Read onTechCrunchMacRumorsPYMNTS

AI content has to be labelled

Article 50 of the European AI ActThe European AI law. It has applied in stages since 2024 and governs what providers and users of AI systems must disclose and check. has applied in full since 2 August. Anyone running a chatbot has to say so. Artificially created or altered images, audio and video need a label that machines can read too. For text the duty falls away if a person reviews it before publication.

The labelling dutyThe duty to mark artificially created content as such. Since 2 August it also applies to companies that merely use AI. applies not only to the providers but to every company using such systems. Fines reach up to 15 million euros or 3 percent of worldwide annual turnover.

Read onBundesnetzagenturLTO

The AI Act now has a German regulator

Also since 2 August the Bundesnetzagentur is responsible for market surveillanceThe state check on whether products on the market follow the rules. The authority acts on its own, without anyone having complained.. It may take samples, request documents, test systems itself and order defects to be fixed.

Start-ups as well as small and medium companies are to get a regulatory sandboxA protected setting in which companies may try out a new application under the authority's supervision. Mistakes are discussed there rather than punished. there. The strict rules for high-risk applicationsAI that decides about people, for instance in hiring, lending or medicine. The strictest requirements of the AI Act apply to it. only take effect at the end of 2027.

Read onBMDSBundesnetzagentur

Google switches off the Assistant on 4 September

Android phones, tablets, watches, headphones as well as Android Auto are affected. Gemini takes over there. Speakers and displays for the home stay unchanged for now.

Gemini is not a successor but something else: The Assistant followed fixed commands, Gemini is a language modelA program that has learnt from a great deal of text how language is built. It follows no fixed commands but produces every answer anew.. Anyone relying on fixed routines should check beforehand whether they still run afterwards.

Read onheise onlineborncity

Washington wants voluntary safety tests

On 4 August OpenAI, Anthropic, Google and Meta met at the White House. Providers are to give the government up to 30 days of access to new models before release.

Nobody has to take part. What gets tested is classified and nothing is published. Open modelsModels whose parts anyone may download and reuse. The maker loses control over who uses them for what. are excluded entirely.

Read onCNBCYahoo News

Deadlines at a glance

• since the labelling duty from Article 50 took effect
• until Google switches off the Assistant
• until the rules for high-risk applications take effect

Recalculated on every visit.

Check in two minutes whether the new duty applies to you

Article 50 has applied since 2 August not only to the providers of the models but to every deployer. A deployer is any company using an AI system, even one it merely bought. On the same day the Bundesnetzagentur took over supervision and may request documents.

The four questions cover what gets overlooked most often. Your answers stay in the browser, nothing goes to a server.