Daily Briefing
23 September 2026 · 5 stories
🔥 Top story
1Researchers discover the first 'malware that is an AI agent' — it makes its own decisions
Cisco Talos published Tuesday the analysis of a brand-new kind of malicious program: a 16-megabyte piece of software that lives inside Windows and, to decide what to do next — steal passwords, hunt for crypto wallets, dig deeper into the machine — no longer waits for orders from a human. It consults a panel of four commercial AIs (DeepSeek, Qwen, Mistral, Gemini), gathers their views, and decides by majority vote. For the general public, imagine a burglar who, instead of waiting for instructions from his boss by phone, asks a panel of experts for advice before every step of a heist — and keeps working even while the boss sleeps. This is the first time such a scheme has been publicly documented on Windows. Source : https://blog.talosintelligence.com/the-closed-quorum-inside-the-first-reported-autonomous-ai-c2-implant/
2Anthropic launches Claude Opus 5.5: as strong as its predecessor, but 40% cheaper to run
Anthropic put into service Tuesday Claude Opus 5.5, the first model of a new generation that matches its previous best on most tasks — but costs 40% less for businesses to run. On per-word pricing, the new model is set at 4 dollars per million words read and 20 dollars per million words written (down from 5 and 25 dollars). On Terminal-Bench, a test where an AI has to write code and fix bugs on its own, it hits 66.4% success — ahead of OpenAI's GPT-6 Astra (57.9%) and Anthropic's previous model (55.8%). For the general public, think of an assistant that used to cost you the price of a coffee to sort through a file: it now does the same job for two-thirds of the price, and does it better. Source : https://anthropic.com/news/claude-opus-5-5
3OpenAI rolls out GPT-6 Sol and Luna: the GPT-6 family is complete, at half the price of the previous generation
A few hours after Anthropic, OpenAI completed its GPT-6 family (already opened with Astra in early September) by releasing two new models. Sol is built for coding and autonomous work: 2 dollars per million words read, 10 dollars per million words written — half the price of the previous model. Luna, a very low-cost version, drops to 0.10 and 0.50 dollars. On a reference test for coding agents, Sol takes first place with a score of 80, while using half as many words, in half the time, and for roughly one-third of the cost. For the general public, it is as if an airline launched at the same time a standard business ticket and an economy ticket costing only a few cents: the most expensive fare stays expensive, but the rest of the cabin suddenly becomes affordable. Source : https://www.testingcatalog.com/openai-launches-faster-cheaper-gpt-6-sol-and-luna/
4Xiaomi open-sources a giant new AI model — with 7,000 training playgrounds for agents
Xiaomi released this Tuesday in open access (MIT license) its new MiMo-V2.6 family, with two models: Pro, which weighs 1,020 billion settings but only uses 42 at a time per question, and Flash, more compact (309 billion total, 15 active). Both handle text, images, sound, and video, and can read the equivalent of a thick novel in a single go. But the biggest piece: Xiaomi also published more than 7,000 training environments where AI agents can practice — code to fix, security flaws to reproduce, websites to build. For the general public, it is like a school opening for free both its textbooks and its sports fields: anyone can now train their own digital assistant with the same tools as the giants. Source : https://mimo.mi.com/docs/news/latest/v2-6
5DeepSeek invited to brief the UN Security Council on AI risks — a first for a Chinese lab
On Wednesday, the United Nations Security Council holds its second formal meeting on the risks of artificial intelligence — the first one dates back to 2023. This time, the labs are at the table: Sam Altman (OpenAI) is travelling in person, Anthropic is sending executives, and DeepSeek, the Chinese lab that shook the planet in January 2025 with its R1 model, is submitting a written text. Its founder Liang Wenfeng is not travelling, but the gesture is historic: for the first time, a Chinese lab is entering the multilateral debate on AI risks at the highest level. For the general public, it is as if Chinese carmakers, who had stayed quiet about safety standards for years, finally agreed to come discuss airbags and seatbelts at the world tribune. Source : https://www.straitstimes.com/asia/east-asia/deepseek-to-brief-un-security-council-on-ai-this-week-sources-say
📡 To watch
Alibaba unveils its own made-in-China AI chip and targets 20 gigawatts of data centers by 2032
The Chinese cloud giant presented Tuesday its new Zhenwu V900 chip, billed as three times more powerful than its predecessor. Mass production is planned for the first quarter of 2027. Alibaba is targeting 20 gigawatts of computing capacity by 2032 — a domestic alternative to Nvidia on Chinese soil. The 'three times more powerful' figure has not yet been tested independently. Source : https://www.bloomberg.com/news/articles/2026-09-22/alibaba-unveils-ai-chip-to-drive-20gw-of-data-centers-by-2032
Snorkel AI raises 350 million dollars to sell 'raw material' to AI labs
This California startup, founded by Stanford researchers, sells ready-to-use training datasets and simulated environments to the major labs, so they can test their agents. Its valuation triples in 17 months to reach 3.5 billion dollars, and its annualized revenue has multiplied by eighteen in a single year. For the general public, it is as if a farm now supplied seeds and test fields to every carmaker: without it, no new model. Source : https://snorkel.ai/blog/snorkel-ai-raises-350m-series-e-at-3-5b-valuation
xAI's Grok 4.7: same price, but faster and better at coding alone for long stretches
xAI made Grok 4.7 available this Monday, at the same price as the previous version (2 dollars per million words read, 6 dollars per million words written). It is already accessible in Cursor and through Grok Build, its own command-line interface. The promise: spend more time on hard tasks and double-check its own work more carefully. The maker claims 'twice as fast at half the price of comparable models' — a figure still to be confirmed on independent tests. Source : https://x.ai/news/grok-4-7
Formal UN Security Council meeting on AI this Wednesday, September 23
It is only the second time the Security Council has met formally on AI risks. Sam Altman (OpenAI) is expected in New York, alongside representatives from Anthropic. On the Chinese side, DeepSeek and Moonshot have been invited to submit written texts — whether their representatives will attend remains to be confirmed. The deliberations could shift the 'AI-pace diplomacy' dials in the weeks to come. Source : https://www.straitstimes.com/asia/east-asia/deepseek-to-brief-un-security-council-on-ai-this-week-sources-say
📊 Trend
This September 22 will be remembered as a turning-point day for AI agents. Three of the biggest Western labs released within a few hours models that are both more capable and up to half cheaper (Anthropic Opus 5.5, OpenAI GPT-6 Sol and Luna), putting enormous pressure on margins across the whole ecosystem. China answered on two fronts: Xiaomi open-sourced a giant model along with its training playgrounds, and Alibaba bet on a 100% domestic chip to break free from Nvidia. On the security side, the most striking news comes from Cisco Talos: for the first time, researchers have documented malicious software that makes its tactical decisions by consulting a panel of commercial AIs — a step beyond the simple flaws in coding tools. Finally, the UN Security Council meets tomorrow, and for the first time a Chinese lab is being invited to speak formally about the risks. For the general public, the signal is clear: AI agents are no longer a demo topic. They have become an issue of international security, trade warfare, and foreign policy.