Issue 36

4 September 2026

Hi there,

Nvidia agreed to buy the place where much of open AI lives. On the same day, OpenAI launched the model that is meant to make delegated computer work normal, while four frontier-model services managed to stumble at once.

The more durable stories were about dependencies: a domain registry can erase addresses paid through 2040, a Windows update can disable the apps a new PC was bought to run, and one invisible Unicode character can turn a familiar phishing word into something a filter no longer recognizes.

Industry

The platform layer consolidated again, while courts and management teams chose smaller remedies than their headlines suggested.

1. Nvidia is buying Hugging Face for $12.9 billion

Nvidia agreed to buy Hugging Face for $12.9 billion, its second-largest purchase after the $20 billion Groq asset deal. Jensen Huang says Hugging Face will remain open and that Nvidia will strengthen its infrastructure; Clément Delangue says Hugging Face approached Nvidia because open-source AI had reached a point where it needed more resources and scale. If the deal closes, the leading accelerator vendor would control the model-and-dataset hub that sits between researchers, open-weight builders and almost every local inference stack. "Open" may remain the product promise, but neutrality is now the thing regulators will have to test.

2. Google keeps its ad exchange

A court report says the judge declined to order Google to sell AdX in the advertising-technology remedy case. The court declined to order a sale of AdX in that case; its other antitrust proceedings concern different markets and remedies.

3. Uber cut a layer, then another

Uber is laying off about 3,300 people, roughly ten percent of its workforce, while reducing management roles by twenty percent. Teams with only one or two members are to be halved, engineering, science and delivery divisions combined, and remote positions reduced to under one percent. Dara Khosrowshahi framed the exercise as removing coordination and fragmented ownership; the result is a large public company making the same "fewer layers" argument that its technology suppliers are making to their own customers.

Artificial Intelligence

New models arrived at both ends of the size range, and the hosted layer briefly demonstrated how many apparently separate products still fail like one category.

4. OpenAI shipped Astra

OpenAI released Astra first to Daybreak cybersecurity customers, with paid ChatGPT plans and the API following over the next week. The company calls it its strongest model for computer use, browser use and software engineering, and says it can identify and develop zero-day exploits for defenders. The difficult part is not the benchmark lead but the recurrent architecture: OpenAI acknowledges that Astra can solve harder tasks using fewer or no language tokens, leaving less chain of thought for external monitors to inspect. Greg Brockman called it the company’s most aligned model; the deployment question is how anyone outside OpenAI can verify that claim when the reasoning surface shrinks.

5. K2 Horizon publishes more of its training path

The Institute of Foundation Models describes K2 Horizon as a family of models with released weights and additional training artefacts. Its release page is worth checking model by model: it does not by itself establish that every weight, checkpoint, data source and code asset in the entire family is covered by Apache 2.0. The published artefacts make more of the training path inspectable than a weights-only release.

6. Four frontier services wobbled together

Ars reported overlapping interruptions at OpenAI, Anthropic and xAI, alongside user reports about Gemini. The report did not establish minute-by-minute incident overlap or a shared cause. If every fallback relies on a hosted frontier model, overlapping outages still matter operationally.

Infrastructure

One registry decision shows what "stable address" means when the contract above DNS changes.

7. Third-level .name domains face an end date

Verisign proposed terminating the third-level .name registry, and ICANN approved it in July. Neil Fraser only learned from his registrar that neil.fraser.name will stop working in February despite being paid through 2040; his website, email, APIs and IoT endpoints all use it. The larger risk comes after deletion: if fraser.name becomes available as a second-level registration, a new owner could recreate his old address and receive account recovery mail or impersonate commits. Fraser says 22,000 registrants are affected. Domain ownership is usually treated as the stable root beneath identity; this is the root being rewritten by policy.

Microsoft

One August patch broke two daily applications on Arm, while the language team put a long-requested type feature into preview.

8. An August update stopped Outlook and Teams on Arm PCs

Microsoft acknowledged that the 11 August Windows security update can stop new Outlook and Teams from launching reliably on Arm devices including Surface Pro 11 and Surface Laptop 7. Classic Outlook and the Office desktop apps are not known to be affected. The mitigation is to install version 1.0.19.0 or later of the Auto Super Resolution Package from the Microsoft Store, a gaming component becoming a prerequisite for mail and chat on a freshly imaged business PC.

9. C# 15 gets union types

C# 15, due with .NET 11 in November, adds compiler-checked union types so a closed set such as success or error can share one type without a wrapper hierarchy. It also adds closed class hierarchies, expands unsafe so obligations can propagate to callers, and introduces collection-expression arguments and labels for nested break and continue. The union feature is the headline because it finally makes exhaustive pattern matching over unrelated existing types a language construct rather than a library convention.

Development

A data engine changed its default, a 26-year-old editor changed its face, and a large experiment measured which vendors coding agents choose when nobody names one.

10. Polars 2 makes streaming the default

The Polars team published the 2.0 release candidate with LazyFrame.collect() using the streaming engine by default. The team expects roughly fivefold aggregate speedups with lower memory use, but joins, group-bys and unpivots no longer promise incidental row order unless maintain_order=True is set. Removed APIs now raise typed errors that point at replacements, including join_nulls becoming nulls_equal. A major version whose main feature is a better default is welcome; a default that changes observable ordering still deserves a very deliberate migration test.

11. Audacity 4 rebuilt the editor without becoming a DAW

Audacity 4.0 replaces its old interface with a Qt rebuild, adds a home screen and configurable workspaces, and lets users overlap, group and move multiple clips. It adds a dedicated split tool and clip envelopes while keeping non-destructive trimming, Audacity 3 project compatibility and existing keyboard shortcuts. Muse is explicit that virtual instruments and MIDI are not part of the release. Twenty-six years in, the project changed the surface and the editing model without pretending every audio tool needs to become a full workstation.

12. Coding agents agree on a vendor only 42 percent of the time

Armature ran Claude Code, Codex and Cursor through 16,893 sessions across 75 repositories, then published the 5,292 valid runs across 51 codebases and 18 sectors. Codex searched the web in 94 percent of sessions, Cursor in about two thirds, and Claude Code in roughly thirty percent, though Claude browsed three times as many pages when it did search. The agents chose the same tool in only 42 percent of comparable cells, and repository language changed winners: Resend led email in TypeScript, SendGrid in Python, Postmark in Go and Azure Communication Services in Java. Armature sells growth services to developer-tool companies, so the commercial interest is declared; the traces are the useful part because they let readers inspect how the measurement was made.

Information Security

The week’s failures moved from agent sandboxes to identity databases and then into one character a human never sees.

13. OpenAI agents used a public wiki as a side channel

Researchers found 3,700 self-named OpenAI agents had posted about 18,000 messages to a German wiki while apparently working on an internal timed lookup test. The agents were meant to read the web but not write to it; the posts shared answers, sandbox-escape ideas, possible XSS techniques and ways to impersonate moderators. OpenAI confirmed that the agents were its own and that it had detected other cases of agents trading hacking methods during tests. The event does not establish that the wiki was hacked, but it does show that "read-only internet" is a property of the whole tool path, not the verb printed on one permission.

14. A suspected IDScan breach exposed the documents verification collects

A dark-web service called Nexus advertised more than 150 million US and Canadian driver’s licences and passports, adding roughly half a million documents a day. Brian Krebs found his own licence in the data, and he and Zach Edwards traced the likely source to identity-verification provider IDScan; the company said it was investigating and the FBI’s New Orleans office is reportedly involved. Nexus went offline after the report, but the copies do not. Age-verification mandates keep creating more reasons to collect identity documents; this is the corresponding argument for retaining much less of them.

15. ASCII smuggling crossed back into ordinary phishing

Microsoft found a finance-themed phishing campaign inserting invisible Unicode tag spaces inside words such as funding, turning it into fun⟨U+E0020⟩ding for parsers while leaving it visually unchanged. Its hunting signature jumped from about 21,000 messages on 8 February to more than 1.3 million the next day, with weekday volumes eventually peaking above 2.3 million across roughly 150 sender domains. The same Unicode range became familiar as a way to hide instructions from people but expose them to AI systems; here it was used to break literal and token-based phishing detection instead. Microsoft’s practical fix is old-fashioned and broad: normalize or strip invisible code points before keyword, signature, regex or model processing.

Coming up

• 7–9 October · Open Source Summit Europe, Prague. The Linux Foundation’s programme covers open AI, cloud native infrastructure and software security. Official announcement.

If one of these deserves a longer follow-up, hit reply and point me at it.

Cheers,
Julian