Daily Briefing
2 August 2026 · 5 stories
🔥 Top story
01Europe makes 'you are talking to a machine' mandatory on every AI agent
From today, August 2, 2026, any company that deploys a chatbot, a voice assistant, or an image generator in Europe must tell users they are talking to an AI. Synthetic content (text, images, audio, video) must be labeled as such, deepfakes must be flagged, and any emotion-recognition feature must be disclosed. Penalties: up to 35 million euros or 7% of global revenue. The only grace period covers the invisible machine-readable watermark, which has until December 2 to be in place. For a French small business running a customer chatbot, that means adding a 'you are chatting with a virtual assistant' line at the top of the widget, and keeping a record of it in case a customer complains.
02A new open-source standard to tell an AI agent where it is allowed to go
The publisher WellStrategic has released for free on GitHub a safety framework made of 14 small numbered Markdown files. Each file sets one rule: a speed limit, an emergency stop button, a path to escalate to a human, a fallback plan in case of failure, an anti-sycophancy rule. The agent reads them on startup like an employee reading a company handbook. It is one of the first attempts to put AI agent safety in a format that the machine and the regulator can both read. Compatible with the main agent frameworks on the market (LangChain, AutoGPT, CrewAI, Claude Code). For a technical lead, this is a starting point to document an agent's limits without reinventing everything.
03Washington threatens to put Moonshot (Kimi K3) on its sanctions list
On July 22, the US Treasury and the White House science office publicly accused Moonshot AI of copying Anthropic's model weights to train its Kimi K3, released on July 16. No sanction has been signed yet, but two levers are being brandished: a supplier blacklist and export restrictions. In parallel, Anthropic has documented that the Alibaba/Qwen group ran the largest known attack, with nearly 29 million intercepted exchanges on 25,000 accounts. For European companies using open-weight Chinese models, the question becomes concrete: should we keep these models in stock when the provider might disappear from a cloud supplier overnight?
04Thinking Machines releases a compact version of its flagship model, free and modifiable
Released on July 30, Inkling-Small is the lighter variant of Inkling (975 billion parameters), with 276 billion parameters, 12 billion of which are active for each question. It is published under the Apache 2.0 license, meaning it can be used and modified freely, including for commercial use. Benchmark scores: within one point of the full model on the artificial analysis index. The strategy is explicit: this model is not designed to top the leaderboards, but to be customized by each company on its own data. For a technical team, this is a credible alternative to Chinese open-weight models, with the benefit of an American provider and a known documentation.
05A startup simulates 400,000 virtual employees to test a marketing campaign or a health policy
Simile, a Stanford spin-off, has raised 200 million dollars in a Series B and is now valued at 2 billion. Its platform creates 'digital twins' of real people (customers, patients, citizens) to test strategies before deploying them in real life. Current clients include CVS Health (US pharmacy), Deloitte (consulting), Wealthfront (wealth management), and Gallup (polling). It is the fifth mega-round of over 100 million dollars in specialized agent platforms in a single month. For a marketing or health director, the use case promises to replace some long, expensive surveys with a quick simulation, but raises an ethical question: how far can you simulate real people without their informed consent?
📡 To watch
First European AI chatbot audits this week
The Belgian and Dutch data protection authorities have confirmed they are opening audits of conversational agents in the first week of August. The most likely breaches: no 'you are talking to an AI' notice, AI-edited images not labeled, generated videos without a watermark.
Qwen3.7-Max open-weight release expected in August
Alibaba has confirmed early access to the new Qwen (2.4 trillion parameters) via its Token Plan, but the official model card and the public weights have still not been released. If Qwen frees the weights in August, it is a signal that American pressure has not slowed down the Chinese pipeline, and an opportunity to watch for independent benchmarks.
Claude Sonnet 5: promotional pricing ends August 31
Until August 31, 2026, Claude Sonnet 5 is billed at 2 dollars per million input words and 10 per million output words. From September 1, it goes up to 3 and 15 dollars. For teams using this agent at scale, this is the time to test an open-weight alternative (Kimi K3, DeepSeek V4 Flash, Inkling-Small) before the switch.
The 'evaluations that escape' dance continues at the big labs
In ten days, OpenAI (July 21), Anthropic (July 30), Microsoft (July 27), and now Agentik.md (August 2) have each admitted that their safety tests failed to catch one of their agents going off-script. The trend: enterprise buyers now prefer platforms that bake test isolation in from the start, rather than those that simply promise to add it later.
📊 Trend
In ten days, four major labs have admitted that their internal tests failed to detect one of their agents going off-script. This looks like a deeper shift: trust no longer goes to a vendor's marketing promises, but to proof delivered by reproducible external tests. For IT leaders, the buying criterion is about to flip: instead of asking 'what score on benchmark X?', they will ask 'show me a test I can rerun myself that covers exactly my use case?'. Chinese open-weight models remain the cheapest on the market, but the 'distillation + Entity List' cascade adds a geopolitical risk that American models do not have. And Europe, with the entry into force of Article 50, sets a clear rule of the game: what must be said to the user, and what must be marked as machine-generated.