The Agent Watch
Briefing · Articles · Tools · About EN FR DE ES 中文 IT PT SV FI DA

Daily Briefing

8 September 2026 · 5 stories

🔥 Top story

1

The Aivres scandal: 5.6 billion dollars worth of Nvidia chips smuggled out of the US into China

Picture a country that bans exporting a technology to another country, then discovers a back door let 5.6 billion dollars worth slip through in two years. That is the Aivres affair, named after the US subsidiary of Chinese firm Inspur, which has been on the American blacklist since 2023. A New York Times investigation reveals that Aivres managed to ship more than 3 billion dollars of Nvidia Blackwell servers to ByteDance and Alibaba, routed through Southeast Asia. For you, this changes the picture: the US-China summit, planned for mid-September, will have to address this hole in AI chip controls. The US Senate is already examining a law, the Remote Access Security Act, to close that loophole. The knock-on effect for users: the rules around Chinese AI models (Qwen, DeepSeek, GLM) could tighten very quickly, and AI chip prices are likely to stay under pressure in the coming months.

2

Anthropic ships ready-made "blueprints" so retailers can build their AI agent before Christmas

Imagine you want to launch a small online store, and instead of hiring a developer for six months, you receive a Lego box with the pieces already assembled: you only have to pick the colours. That is exactly what Anthropic has just put online for big brands: a library of ready-made 'plans' to build shopping, basket, and after-sales conversational agents in a few days instead of several months. The timing is strategic: the American high shopping season (Cyber Monday) lands on December 1st, and Fortune 500 retailers no longer have time to rebuild everything from scratch. For an SME selling online, this could mean it will soon be able to plug this kind of agent into its own site without hiring a specialised technician.

3

GitHub Copilot becomes a team of specialised agents, and OpenHands 1.0 ships as a stable release

Picture a site foreman who, instead of doing everything himself, hires a plumber, an electrician, and a bricklayer — each on their own specialty, but all sharing the same blueprint. That is what GitHub Copilot now allows: you launch a coding mission, and several specialised agents (one for implementation, one for tests, one for documentation) collaborate while sharing the same context. In the same move, OpenHands 1.0, an open-source coding agent, reaches stable release and claims roughly 68% success on a recognised programming benchmark. For an SME, this is a game-changer: you can now hand a real coding mission to an agent for a full day, without having to babysit it at every step.

4

A New York start-up raises 50 million to build the "gatekeeper" of enterprise AI agents

Imagine giving a new employee the keys to your office before leaving them alone: you first check that they will not open the wrong drawers. That is exactly what Air, a young New York company founded in January 2026, offers. Its 50 million dollars raised from Sequoia and Greenoaks is being used to build a 'gatekeeper' that checks every website, every extension, and every model before an AI agent connects to it. Four steps: static analysis of the site, dependency check, sandbox testing, and a trust score updated continuously. For a company hesitant to deploy agents on sensitive data, this is the missing link — and it now sells at a very high price, opening a brand new security market.

5

The first smartphone that "forgets everything" arrives: an on-device AI agent, no cloud

Imagine a phone that remembers everything you tell it… unless you decide otherwise. That is the bet behind the Nubia NaviX Ultra, launching on September 16 in China: a smartphone with an AI agent baked into the system, an orange physical button to wake it, and a 'Global Memory' that stays on the device by default. No data leaves for the cloud without your consent, and a factory reset wipes everything. For European users wary of their data, this is an important signal: AI is starting to become an object you hold in your hand, not just a distant service you talk to. The stakes are also strategic: if this model takes off, Apple and Google will probably have to offer a comparable option.

📡 To watch

The RASA law on remote security of AI chips

The US Senate must vote on the Remote Access Security Act, already adopted by the House. If it passes, controls will extend to remote access via the cloud — forcing AWS, Azure, and GCP to tighten their export conditions for models and compute.

The IPO of Nscale

The London-based AI cloud provider Nscale is preparing its US IPO with a 30 billion dollar valuation and a 103 billion dollar order book, boosted by a 45 billion dollar, six-year contract with Anthropic. To watch as soon as the official announcement is made, expected this September.

Anthropic's enterprise safeguards

On September 2, Anthropic announced its Enterprise Frontier Safeguards framework, delivered 'in the autumn of 2026': zero data retention and detection of misuse for regulated customers. To verify as soon as it goes live.

Adoption of the Nubia NaviX Ultra smartphone in Q4

If Chinese consumers take to this 'privacy-first' model, giants Apple and Google will be forced to integrate a comparable on-device AI agent option. First real-world test right at launch, on September 16.

📊 Trend

Three forces are crossing paths this morning. First, agent security is becoming a real market: the start-up Air raises 50 million dollars to verify websites before agents connect to them, and the scandal of Nvidia chips smuggled into China is pushing Washington to tighten the law. Second, hardware is becoming the agent: the Nubia NaviX Ultra is the first smartphone with an AI agent built into the system and a memory that stays on the device. Third, agents are moving from minute-by-minute assistant status to coworker status: GitHub Copilot now has several specialised agents collaborating on the same code, and Anthropic ships ready-made plans so retailers can launch their agent before the holidays. For a company, the lesson is clear: agents are arriving in production — but they arrive with a security contract to sign and a hardware choice to make.