Daily Briefing
21 September 2026 · 5 stories
🔥 Top story
1A major flaw discovered in the main 'agents that code for you'
Picture a personal assistant you hand your keys to: this week, researchers showed that four of the biggest AI coding assistants — those from Anthropic, OpenAI, Google, and Microsoft — could be forced to run malicious code without the user even clicking. It is the first attack of its kind targeting the supply chain of extensions for these tools. Three vendors have already shipped a fix; Microsoft has still not released a patch. For the general public, it is a bit like if the lock on your smart front door could be opened without breaking in by anyone who knows the trick. Source: https://www.theregister.com/2026/09/18/plugin4shell/
2A US military chatbot 'almost started a war' with China
Last spring, a US military analyst asked an AI tool to review the cargo manifest of a Chinese ship in the Mediterranean. The software concluded, out of thin air, that the vessel was carrying nuclear material. Fighter jets took off, special forces were ready to board — before someone dug out the report and realised everything was wrong. For someone watching from the outside, it is a bit like a GPS sending you into a lake while swearing it is the highway. Source: https://www.cnn.com/2026/09/18/ai-military-hallucination/
3Trump calls AI safety a 'hoax' and creates an 'AI Force'
On Friday, the US president announced the creation of an AI-focused force modelled on the Space Force, and appointed an 'AI czar' to speed up the technology race. At the same time, he dismissed as a hoax the calls from tech bosses (Anthropic, OpenAI, Google) asking to slow down the development of the most powerful models. In the same days, the governor of California signed the opposite: an executive order imposing an emergency stop button and outside auditors for frontier models. For the general public, it is like two captains on the same boat pulling the wheel in opposite directions. Source: https://www.nbcnews.com/2026/09/19/trump-ai-force-czar
4China releases a multimodal AI model seven times cheaper than GPT-5
On Saturday, the Chinese lab StepFun opened a new model that can read text, images, and video, and can remember around a million words worth of conversation at once. The announced price is about one dollar per million words read, and 2.70 dollars per million words written — that is, at equal power, roughly seven times cheaper than OpenAI's flagship model. For someone who does not follow the details, it is a bit like a budget airline offering the same seat as the national carrier, with a ticket at one seventh the price. The model weights will be opened on October 15. Source: https://stepfun.com/step-5-preview
5A 706,000-parameter mini-model to fill in your forms for you
On Saturday, a young French startup released a tiny AI model specialised in a single task: filling in forms on screen. Seven hundred and six thousand parameters, barely three megabytes, trained in a single pass — that is, without generating text token by token. For the general public, it is the equivalent of a hyper-specialised employee who only knows how to do one thing, but does it very fast and for almost nothing. It is the first of a planned family aimed at high-volume repetitive tasks. Source: https://buildfastwithai.com/blog/cua-s1-forms
📡 To watch
US and Chinese finance chiefs meet in New York before the Trump-Xi summit
On Sunday, US Treasury Secretary Scott Bessent and his Chinese counterpart He Lifeng spent the whole day at JPMorgan Chase's Manhattan headquarters preparing the Trump-Xi summit on 24 September. AI is explicitly on the official agenda. Beijing is rejecting the US calls for a slowdown as 'fear-mongering'. Source: https://www.wionews.com/2026/09/21/bessent-he-lifeng-meeting
Microsoft Copilot: still no fix for the extensions flaw
Among the four main coding assistants hit by the flaw discovered this week, only Microsoft's tool has still not received a patch. About 90% of the Fortune 500 use Copilot, according to Microsoft. The risk remains open. Source: https://www.theregister.com/2026/09/18/plugin4shell/
Alibaba releases a multimodal model and changes its open source licence
On Friday, the Chinese giant Alibaba released Qwen3.8-Omni-Flash, which can handle text, image, audio, and video, along with Qwen-Image-2.1 to generate images. But it simultaneously moved from a very permissive open licence to an in-house licence that bans commercial use without an agreement. The thousands of products using the old framework must renegotiate. Source: https://qwen.ai/blog/qwen3-8-omni-flash
The Chinese Step 5 model will open its weights to the public on October 15
StepFun's new multimodal model is already available through a programming interface, but its weights — the downloadable core of the model — will only be released on October 15. For the open source community, this is the event to watch: a frontier model at a rock-bottom price, in the open. Source: https://stepfun.com/step-5-preview
📊 Trend
This weekend pushed the topic of AI from the technical field into the political one: for the first time, a US state (California) took a stricter position than the US president on the safety of frontier AI models, while the president was launching an 'AI Force' and calling safety a 'hoax'. At the same time, a new zero-click flaw was discovered in the four main coding assistants, proving that AI tools have become a standalone attack target. On the Chinese side, a new multimodal model landed at one seventh the price of equivalent US models, and a major Chinese house (Alibaba) tightened its open source licence. For the general public, the signal is clear: the AI race is no longer just fought in the labs, but also in courts, diplomatic summits, and licence terms.