Editor's note
Today was about the governance and the plumbing of AI. The UN Security Council convened its first high-level briefing on AI under France's presidency, seating US and Chinese frontier developers together. OpenAI said it will let outside evaluators test its models from the training phase, and a group of banks warned that AI shopping agents are outrunning fraud protections. On the product side, xAI shipped Grok 4.7 and Alphabet's Intrinsic open-sourced its robotics control stack under Apache 2.0. The through-line was less about flashy launches than about how autonomous, agentic AI should be overseen and what parts it will run on. The Web side was quiet, so today's issue leans toward AI.
AISecurity Council Report
UN Security Council holds first high-level AI briefing, seating US and Chinese developers together
On September 23, the UN Security Council holds a high-level briefing on AI and international security. France, the Council's president for September, convened it, with Foreign Minister Jean-Noël Barrot chairing. Briefers include Yoshua Bengio (co-chair of the UN scientific panel on AI), OpenAI's Sam Altman, Anthropic CEO Dario Amodei, and Hugging Face CEO Clément Delangue, alongside invited Chinese developers such as DeepSeek and Moonshot.
Context
France's concept note frames the session around two worries — malicious uses of AI and the risk of humans losing control of increasingly capable systems. It is the first time the Council has seated frontier AI developers from both the US and China in the same room, framed as a dialogue on shared safety concerns rather than a regulatory forum. This is a stage for aligning perceptions, not a binding decision; expect statements of direction more than concrete rules.
HaLVision's take
What draws our attention is less any single vendor's message than the fact that several parties are in the room together. The direction voiced here will eventually bear, indirectly, on the terms and data handling of the AI that small businesses use. Rather than predicting which country's framework prevails, we advise keeping a design where you control how data is held and scoped. The likely output is a declaration, so temper expectations.
AIOpenAI
OpenAI to let outside evaluators test its models from the training phase
On September 22, OpenAI said it will let third-party groups assess its models for safety risks earlier in the development cycle. Rather than only near release, it intends to allow technical safety assessments during training and evaluation. It set out principles for good reviews — strong independence, scientific rigor, robust security, and clear responsibilities — and said it is in talks with groups including METR and Redwood Research.
Context
The move responds to a persistent worry that safety checks lag capability gains. Bringing outside eyes in during training can surface problems earlier, but it raises questions about protecting sensitive material and preserving evaluators' independence. For now no partner or access terms are settled, and the company frames it as a direction — with the most sensitive work possibly done by evaluators brought on-site.
HaLVision's take
Widening the window in which an outside party can inspect a model is, on transparency grounds, welcome for users. That said, read it with the discount due any vendor's voluntary program; its real weight rests on how independence and terms get nailed down. For those of us embedding AI into client work, how far a provider opens itself to third-party review is a reasonable factor to weigh at selection time.
AIxAI
xAI ships Grok 4.7 — stronger reasoning at the same price, pitched for coding
On September 21, xAI released Grok 4.7, calling it an improvement over Grok 4.6 at the same price and speed. It carries roughly 2.1 trillion parameters (up about 40% from 4.6's 1.5 trillion), spends longer on hard problems, and double-checks its own answers more often. Pricing is $2 per million input tokens and $6 per million output, and it is available in Cursor, the Grok API, and GitHub Copilot.
Context
Grok 4.7 arrives after several delays and leans hard on coding performance. xAI folded in supplemental data from SpaceX — Starlink telemetry and manufacturing and failure logs — and pitches stronger reasoning about physical, hardware systems. It claims benchmark gains, though its timing puts it somewhat behind the frontier pack this cycle.
HaLVision's take
What we watch in practice is less the marketing than price-performance and how cleanly a model drops into existing tools. Another same-price option selectable from GitHub Copilot or Cursor widens the choices for code assistance. The sound move is to benchmark it head-to-head on your own routine tasks — scaffolding, refactors — and decide on the cost-quality balance. We recommend a setup you can switch out, rather than locking to one vendor.
AIQuartz
Bank coalition warns AI shopping agents are outrunning fraud protections
On September 22, a coalition including Bank of America, NatWest, ING, New Zealand's ASB Bank, and Capital One published a joint report warning that agentic commerce — AI selecting and buying goods — is creating new fraud, scam, and data-privacy risks. It says the rules meant to protect shoppers have arrived slower than the tools, laid out development principles, and said it plans to bring proposals to policymakers.
Context
A central worry is that some agents type a shopper's card details straight into checkout, skipping the step where a person would catch an error or a red flag. The report also flags that agents could quietly route purchases through payment options carrying weaker fraud protection. Coming a day after Amazon blocked Meta's shopping agent, it shows the financial side now pressing for lines around autonomous buying.
HaLVision's take
This lands directly on anyone running an online store or site. Soon you will decide which agents may go as far as checkout, and within what limits. The safe path is to treat unidentified automated agents like ordinary bots and to settle, in advance, your policy on identity checks, agent identification, and rate limits at the payment step. Before agent-driven purchases grow, design where a human confirmation stays in your order flow.
AIIntrinsic
Alphabet's Intrinsic open-sources its robotics control stack under Apache 2.0
On September 22, at ROSCon 2026 in Toronto, Alphabet-owned Intrinsic unveiled Intrinsic Core, releasing the core of its industrial robotics platform under an Apache 2.0 license. It is ROS-compatible and includes real-time control, motion and grasp planning, machine vision, simulation, and hardware integration. It is available on GitHub, alongside a reference design for CNC machine tending.
Context
Its notable parts are a hardware-agnostic real-time control framework that adjusts a robot's path mid-move from sensor feedback, and pose estimation built on Nvidia's FoundationPose. By standardizing infrastructure firms used to build from scratch, it aims to widen the base of so-called physical AI development. A large company opening core parts under a permissive license also serves to draw in a developer community.
HaLVision's take
Our main line is web work, but we read this kind of opening as a barometer for the going terms of AI adoption. A big vendor putting core parts out under Apache 2.0 signals a judgment that growing the base beats fencing it off. The same dynamic — open the platform, win on the ecosystem — plays out in web and app tooling too. When picking the tools you rely on, it is sound to weigh license and the freedom to switch.