Key points
- Java4Shell — Security researcher Calif published a technical disclosure for the Java4Shell vulnerability, tracked as CVE-2026-114938.
- Quad9 — The free non-profit public DNS resolver Quad9 announced that, from 24 November 2026, it will stop providing its free network protection to users in France.
Here is your Daily AI Stew, served up by Robin.
Microsoft moves Office into Copilot, and Nadella calls it the new OS for work
We start with the biggest company of the day, and the point is not the model but the editor you open every morning.
Opening the Ignite keynote, Microsoft CEO Satya Nadella laid out the company’s vision for a new “frontier ecosystem” and announced that Word, Excel and PowerPoint are already available inside Copilot, which he called “the new operating system for work”.
Ars Technica describes the launch as a reversal of roles: Word and its peers used to star, with the assistant helping on the side like Clippy, while the new Copilot integration is driven by agents and the three Office apps ride along as plug-in parts, backed by an organisational knowledge layer that pulls context from mail, meetings, calendars, chat and files. Users can now create documents, spreadsheets and presentations directly inside Copilot, and the generated files can still be edited in the corresponding Office apps.
The real change is where the entry point sits: apps used to be the subject and the assistant an attachment, and now it is the other way round — apps become components that get called. That favours Microsoft, which sidesteps the incumbent fight at the application layer and stands at the starting point instead; for users it is another bill, because the files are still there but habits are being asked to move, and whether that organisational knowledge layer really ties mail, meetings and calendars into context decides whether this turns out to be just another sidebar that gets switched off. Next to watch is the permission model: who can read which mail, and who answers when something goes wrong — until that is spelled out, enterprises will not hand over their real documents.
Java4Shell: an OpenJDK Base64 decoding bug that can lead to remote code execution
Over to security, where this one hides in a decoder almost every Java service touches.
Security researcher Calif published a technical disclosure for the Java4Shell vulnerability, tracked as CVE-2026-114938. The flaw sits in OpenJDK’s Base64 decoder: decoding can write past the bounds of the target array, which creates the conditions for remote code execution.
Per the disclosure, this exploitation path needs heap-layout manipulation over the network — so-called “Remote Heap Feng Shui” — and its reliability depends heavily on round-trip latency and jitter, with low, stable latency helping the attacker; the researchers completed a real attack over a public IP in a controlled environment. OpenJDK and Java SE deployments on affected branches face heap corruption, disclosure of secrets and remote code execution from an unauthenticated attacker.
The disclosure also carries detection advice and mitigation guidance. Semgrep has released a free detection script, and one of the five conditions it lists is that only HotSpot generates the relevant stub, so OpenJ9 and GCJ are unaffected. The publicly known exploitation method is demanding about network conditions and JVM environment, though, and real exploitability depends on the specific deployment and network — it cannot be treated as “every Java app is now compromised”.
What makes this kind of bug frightening is not how clever the exploit is but how unevenly the patching is spread: Java’s long tail often sits inside someone else’s product, where the vendor sets the upgrade cadence and operators may not even know which branch they are running. The disclosing side is right to ship a detection script, but detection is not repair, and those five conditions work more like a sieve — every hit still needs a human to confirm it. What to watch next is when each distribution pushes the patch down, and what happens to branches nobody maintains any more; that part usually goes unowned, and that is exactly where things break.
Nvidia’s third-quarter revenue rose 103% to $115.8bn, and the board says Huang will not return
Now to the capital markets, where a very pretty earnings release hides a governance question.
According to Yahoo Finance, Nvidia’s third-quarter revenue rose 103% year on year to $115.8bn, beating expectations and lifting guidance; data-centre revenue rose 110% to $107.6bn.
It is Nvidia’s first quarterly report since founder and long-serving chief executive Jensen Huang left. The board said it will not reinstate Huang as chief executive, and interim CEO Cress continues to run the company.
The report also notes that Chinese cloud providers and intermediaries renting Nvidia’s advanced chips through brokers are drawing more compliance scrutiny, because Huang’s departure has put the related Asian partnerships — and the possibility of using brokers to sidestep US export limits on advanced chips — under closer review.
The growth curve while the founder was in charge and the first scorecard after he left are almost equally steep, which by itself weakens the “irreplaceable figurehead” narrative; by ruling out a return at this particular moment, the board is moving the company’s value from a person to an organisation. The real test is not this quarter but the roadmaps and long-cycle commitments of the next few, and an interim CEO usually lacks the authority to sign those. Another variable is compliance review on the channel side, where tightening usually hits the flexible but marginal demands first, not the big customers.
Microsoft ships MAI Code 2 and Code 2 Flash, available to paying GitHub Copilot users
Back to tooling, where Microsoft has taken another step with its own models.
Microsoft announced MAI Code 2 and MAI Code 2 Flash, two code-generation models from its in-house AI team, aimed at AI-assisted coding in software engineering. In Microsoft’s framing, Code 2 is the workhorse that handles coding tasks and backs Microsoft products, while Code 2 Flash is light, low-cost and suited to fast iteration.
In its 17 November Changelog, GitHub lists code explanation, targeted edits and routine fixes as typical uses for Code 2 Flash, which says the pair targets high-frequency development workflows rather than generic code generation. Eligible, annually billed GitHub Copilot users can pick both models now, across Free, Student, Pro, Pro+, Max, Business and Enterprise tiers.
Microsoft says Code 2 beat GPT-6 in 64% of blind tests on front-end applications, but its behaviour in real workflows still needs checking. The available material gives no model architecture, training method, benchmark results, pricing, licence, availability regions or tool compatibility, so the actual capability gap between the two cannot be judged yet.
The point of in-house models is not benchmark scores but bargaining power: as long as coding capability depends on outside suppliers, Microsoft has to trade off cost against capability on someone else’s terms. Shipping the lightweight version first to carry high-frequency tasks is pragmatic, because explaining code and pinpoint fixes are sensitive to latency and price and not to the ceiling. Curiously, the only number officials gave is that blind-test preference, with nothing on architecture, pricing or licensing; that restraint suggests the real gap is not yet ready to be discussed openly. What to watch is not the leaderboard but whether anyone dares to admit publicly how completion quality changed once the default tier was swapped.
Quad9 stops offering free network protection in France from 24 November
One more change that is less visible but very wide in reach: a free public utility is leaving a market.
The free non-profit public DNS resolver Quad9 announced that, from 24 November 2026, it will stop providing its free network protection to users in France. The reason is a French court ordering it to block sites accused of copyright infringement, and under that legal obligation it chose to withdraw. Quad9 resolves domain names to IP addresses and adds an extra layer of security on top.
Cloudflare’s blog reads the event as one more case of the tension between national regulation, public DNS operation and the governance of open internet infrastructure. With Quad9 gone, French users have one public DNS option fewer and need to turn to other privacy-friendly resolvers or to the resolution service their network operator provides; other cross-border public DNS operators may also adjust per jurisdiction and exit specific markets when compliance requirements conflict.
Quad9’s chief technology officer said the IETF 127 meeting showed 100% support for its position at the time, while stressing that this was only a point-in-time consensus. Some commenters argue the French case is not necessarily a canary in the coal mine and may simply be an example of a national government successfully implementing internet regulation.
The value of these services is invisible most of the time, and users only notice after they are withdrawn that they were relying on a non-profit’s goodwill. An operator stepping away in the face of a legal demand shows that a small organisation cannot carry that conflict cost long-term and can only leave. For developers, the choice of resolver becomes a local variable again: the same configuration may stop working in another country. Worth watching is whether other cross-border resolvers take the same road, and whether operators’ built-in resolvers can cover the privacy side.
Solid 2.0 ships: server RPC, live broadcasts and diagnostics written for coding agents
Framework news normally stays inside front-end circles, but this one moves something that matters.
SolidJS released 2.0, a major version whose focus is pulling server RPC, live/broadcast realtime capability and diagnostics aimed at developers and coding agents into one place. Solid is a JavaScript framework built on fine-grained reactivity, with Context and Store providing built-in state management so global state does not need a third-party library.
To shrink the layer between front end and back end, Solid 2.0 moves capabilities that meta-frameworks used to carry — server functions, single-flight mutations, filesystem routing and deployment handlers — into the 2.0 core or into framework-agnostic Vite plugins, and server RPC and live/broadcast follow that same line.
Community response leans positive: pwmglenn finds it more natural and coherent than LiveView and htmx/Datastar, pier25 calls it a milestone release, and wk_end argues that “writing for agents” is really better debugging and diagnostics, since the best practices humans follow apply to agents too. Others note Solid’s component library is still thin and are waiting for the performance-led Vue 3.6 Vapor. Two caveats: at the time of writing the Solid site still labels 2.0 a release candidate, and no complete release notes exist, so the specific APIs, compatibility, benchmark data and migration requirements cannot be confirmed; there is also a developer write-up of a dashboard A/B test reporting a 40% interactivity gain over four weeks and 10,000 user sessions, but that is a single case.
Blurring the boundary between framework and meta-framework pays off immediately in simpler project structure and risks a narrower upgrade path in the long run: two layers that could be replaced separately now have to move together. Pulling server functions and routing into the core amounts to admitting that the old split cost too much in practice, and that self-correction deserves credit, but what the community really needs is migration documentation, which is exactly what is missing. The site still marking it a release candidate is a detail that argues for patience about the release cadence. What to watch is whether it can make the realtime side solid without giving up what fine-grained reactivity is good at.
The “All About Berlin” author says AI summaries are hitting traffic, and next year full-time may not hold
One last item from an individual’s viewpoint, about the foundation the content business stands on.
Nicolas Bouliane, who owns the independent Berlin living guide “All About Berlin”, published a first-person piece attributing the site’s falling reach and the pressure on his livelihood to AI’s effect on original online content and its business model. In the Hacker News discussion he added that conditions kept worsening months after publication, that he can still support himself from the site, and that next year he may not be able to run it full-time.
All About Berlin is an independent Berlin guide Bouliane has run since 2017. After Google’s AI Overviews launch, the traffic curve he published turned from years of growth into a steep fall; the “zero-click web” means users get a summary answer on the search page and no longer click through to the pages being cited. The core tension in the discussion is that generative AI depends on pages humans write while potentially undermining the economics of those sources by cutting visits and referrals.
There are other views. Some commenters point out that Google never promised free traffic to begin with; others argue users may prefer a summary assembled across sites over links to sources of uneven quality, potentially outdated, and shaped by affiliate commissions. Nicolas also mentions receiving a flood of business email generated by large models. The available material is mostly personal experience, with no quantitative data on traffic or revenue and no systematic study, so it cannot show that AI is the sole cause of the site’s decline.
A site run by one person eventually gets squeezed to the edge of viability by a few lines of summary on a search page, and that ending is blunter than any industry report. There is no conspiracy here, only misaligned incentives: models can only be trained on what people write, and those same people’s readers are being kept on the results page. He can hold on personally, but for independent writers like him exiting is a matter of time, and content supply thins first and then becomes a public good a few people are willing to maintain for free. The question is not whether he has the right to complain, but who will still pay for first-hand reporting that requires going somewhere once the retrieval layer becomes the answer layer.
Put together, today’s stories point the same way: the entry point is folding into each platform’s own shell, and the security debt and the content business’s bills are still left for developers to pay. That is all for today.
🎧 This episode is also available as a podcast: listen to Daily AI Stew · 2026-11-18.
