Alibaba Cloud will open its first Dutch and Finnish regions within a year
Alibaba Cloud said on 23 September that over the next twelve months it will establish its first cloud regions in Türkiye, Finland and the Netherlands, and expand its data centre footprint in Germany, France, Malaysia, the UAE and Hong Kong. The figures in its own announcement are 107 availability zones across 31 regions today. No opening dates, capacities or investment amounts are given for any of the new sites; CTO Feifei Li told the Apsara Conference in Hangzhou that the Dutch region comes first, in October, as SCMP reported, but that date is not in the written release.
For a European buyer the useful question is not latency but law. A region inside the EU gives data residency; it does not change who can be compelled to hand data over, and a hyperscaler headquartered in the People's Republic of China sits in a different risk class from a US-incorporated one rather than a safer one. The same announcement launches Smart Fusion, a routing layer that picks between models within one task behind a single endpoint and claims roughly 50% lower token spend, and Smart Studio, which claims up to 505% higher throughput than standard open-source serving frameworks on the same hardware. Both numbers are Alibaba's own, with no published methodology and no baseline named.
Black Forest Labs ships a robotics model most European manufacturers cannot deploy
The Freiburg lab published FLUX 3 Action on 23 September: a 7B open-weights world action model that takes a camera frame and an instruction and returns the next two seconds of robot actions. Its own results table puts it first on RoboLab-120 at 42.92% success, ahead of OASIS WAM on 39.0% and NVIDIA's 16B Cosmos3-Nano-Policy on 36.8% — a win at under half the parameters of the open-weights runner-up. The score is the lab's own, published on its own page, on a benchmark built by NVIDIA; the table's caption says only that model types and weight availability follow NVIDIA's leaderboard, not the success rates. Treat it as a vendor claim until the leaderboard shows it. Checkpoints are integrated into LeRobot and the code is on GitHub.
The licence is the part to read. Unlike the tag-without-a-file pattern that has run through open-weight releases all month, a real LICENSE.md exists — and it is the FLUX Kommunity License v1.0, which grants commercial and production use only to a “Qualifying User”, defined as one whose gross annualised revenue together with its affiliates is under US$5,000,000. Everything above that line is non-commercial and non-production only, and the licence defines “Robotics Uses” expressly to cover any production robotic system or embodied AI system. The lab's own launch post nonetheless lists FLUX 3 Action under “Open source: Yes”. Any industrial firm large enough to run a robot cell needs a separate paid licence, which makes this a commercial negotiation dressed as a download.
An independent benchmark put a frontier model in a real car, and got past its refusal by renaming a tool
Three researchers built DrivingBench, wired a laptop running an agent loop into an openpilot-equipped car over a phone hotspot, and published the full report with the harness open-sourced and every trace released. GPT-6 Astra was the only model to complete the physical course; GPT-5.6 Sol, Grok 4.6 and Claude Fable 5.1 all failed. It covered 134.7 metres in 5 minutes 22 seconds — about 0.94 mph — at $7.74 in tokens, which works out to roughly $92 per mile against about $0.18 for a 25-mpg car at $4.60 a gallon. Each model got one evaluation at one reasoning effort, which the authors say plainly; treat the ranking as a single observation, not a league table.
The finding that travels beyond cars is in the methodology. Some models — Astra especially — would sometimes refuse to drive, and after a failed attempt at calling the run a “simulation”, what worked was renaming the MCP server to “DrivingBench Sandbox”. That is a relabel of the surrounding context, with no change to the capability being exercised. Anyone relying on model refusal as a control in an agent deployment should read it as a documented bypass, and note the token arithmetic while they are there: the blended rate came out near $1.17 per million against a $10 list input price, because almost everything was billed as cached input at $1 per million. The Register covered the run, with one error worth not repeating — it has xAI hosting GPT-6 Astra, which is OpenAI's model.
Quick Hits
- Spain's regulator warned a company about CV-screening AI it had not yet switched on. — The AEPD published a formal warning on 23 September to a firm evaluating a tool that would screen applications, assign scores and prioritise candidates for hiring and internal mobility. The system was not yet live, though staff had been told it was coming. This is an advertencia under Article 58(2)(a) GDPR — a preventive measure, explicitly not a finding of infringement, and not an AI Act action. What it asks for is data protection by design, a prior impact assessment where risk is high, and transparency about the score. Employment-screening AI is exposed on the data-protection track today; its Annex III high-risk obligations apply from 2 December 2027.
- An LLM router shipped an authentication bypass via a header the client controls. — GHSA-5mj8-gf6m-fhw8 (CVE-2026-56681, High, 7.3) describes 9router treating any request carrying the header X-9r-Real-Ip: 127.0.0.1 as loopback and skipping API-key validation on its public /api/v1/* surface — free use of the owner's provider credits, on deployments that do not run the bundled custom-server.js wrapper. A companion advisory (CVE-2026-56682, Moderate, 5.3) shows the login rate limiter bucketing on the same spoofable header, so rotating it gives unlimited guesses against a dashboard with a default password. Both fixed in 0.5.8. Worth grepping your own gateway for trust inferred from a proxy header.
- A self-hosted RAG stack got five advisories, and one of them fires from your own documents. — Published to the GitHub database on 22 September: CVE-2026-85734 (Critical, 9.1) is a POST /login with no rate limit, lockout, backoff or CAPTCHA. Alongside it LightRAG carries an SSRF bypass using IPv6-transition addresses, a stored XSS in the WebUI answer renderer triggered by ingested content, non-constant-time password comparison, and raw exception messages in API responses. Patched in 1.5.5. A poisoned corpus reaching admin operations is the realistic chain.
- The MCP TypeScript SDK now caps request bodies by default. — Version 1.30.1, released 23 September, caps HTTP bodies at 4 MiB by default — configurable via maxRequestBodySize — and rejects JSON-RPC batches over 100 messages with a 400; previously an unbounded batch was a cheap denial of service. The 2.x packages apply the same ceiling across every read path the SDK owns, and make Streamable HTTP transports return HTTP 403 insufficient_scope before a handler executes when a registered primitive carries a scope challenge. Large tool payloads that worked yesterday will be rejected after the upgrade.