AI & Tech Daily
Office moves into Copilot, Java patches land, and Gemini splits privacy controls
Microsoft brings Word, Excel and PowerPoint into Copilot, opening a practical test of whether integrated document work reduces hand-offs and rework. Reson8 releases downloadable multilingual transcription weights; Java gets a hardware-specific security fix; and newly public US government tests document chatbot safeguard failures. NVIDIA outlines a scientific-computing GPU for 2028, while Australia joins an AI supply-chain declaration. Plus, what Hugging Face's routing and billing change means for developers, and Gemini's new separation of chat history from training consent.
Full transcript
Read the episode.
I'm Jesse Owen. This is AI and Tech Daily.
Office moves inside Copilot
An AI draft is only part of getting a document finished. Microsoft is bringing Word, Excel and PowerPoint inside Copilot, putting the editing and collaboration in the same place as the conversation.
Our main story today is Microsoft's move to bring Office work into Copilot: can keeping the conversation and the file together reduce the work between a first draft and something colleagues can use?
At Ignite on 17 November, Microsoft made Office in Copilot and its Home interface generally available. People can create and coauthor Office files within Copilot, alongside chat and delegated tasks. For organisations with the relevant Copilot subscription, that creates a more complete place to work on a document.
The interesting part is what happens after the first answer. A useful proposal may need changes from its author, a colleague's contribution and another round of editing. If those steps can happen around the Office file in the same workspace, there's an opportunity to spend less time transferring material between a chat window and the document people are reviewing.
That's the workflow I'd want to examine. Consider a manager preparing a proposal with a colleague. The first draft might arrive quickly, but the time-consuming part could be making the argument precise, checking that the figures support it and incorporating the colleague's edits. Integrated drafting and coauthoring give that work a shared home. Whether they shorten the whole job depends on how well the editing and collaboration fit the way those people work.
Microsoft says the files retain Microsoft 365 permissions, sensitivity labels and compliance controls. Those are meaningful details for an organisation already using Office. Permissions govern who can access a file; sensitivity labels carry its information classification. Retaining those controls means the document workflow can be assessed within the organisation's existing arrangements for handling information.
It also gives the people evaluating Copilot a concrete object to follow through the process. They can look at the document colleagues end up using, the corrections it needed and the time spent getting it there. A fast opening draft is useful. The result becomes more compelling if less effort is needed to turn that draft into an agreed piece of work.
There are different release stages within this announcement. Office integration and Home are generally available. Autopilot and Workspaces are beginning their Frontier early-access rollout. Copilot Code is due in the coming weeks. Those dates let an organisation separate the document capabilities it can evaluate now from the autonomous features still arriving.
The charging model deserves the same practical treatment. Office integration is included in the Copilot subscription licence. Other work can attract consumption charges, so the cost of a workflow depends on which capabilities it uses. A subscription price on its own may leave part of that calculation out.
For a buyer, I'd compare the cost and effort of completing a familiar document task, including the edits and collaboration. How many hand-offs disappear? How much rework remains? Does the finished file arrive sooner, with less effort from the people responsible for it? Those questions connect the software purchase to a result the organisation can recognise.
Independent productivity measurements for this new setup aren't established yet. The opportunity is clear enough to evaluate: keeping more of the document's working life beside the AI assistance could make that assistance useful beyond the first draft. For organisations already paying for Copilot, the newly available Office workflow is the immediate place to test that proposition.
Downloadable multilingual transcription
That's the big workplace shift. A smaller release gives developers more control over where another everyday task happens: transcription.
On 18 November, Reson8 released downloadable weights for Rumble-2-Flash, its speech-recognition model, under the Apache 2.0 licence. The model had already been available through an API launched on 3 November. Developers can now run the model themselves.
Rumble-2-Flash has half a billion parameters, supports more than forty languages and integrates with Hugging Face Transformers. Parameters are the learned values within the model; the relatively small size makes this an interesting option for developers weighing the resources needed to operate their own transcription service.
The downloadable weights and permissive licence change the deployment choice. A builder can supply the hardware and runtime, then integrate speech recognition into an application without making a hosted transcription API compulsory. That opens up more control over the processing environment and how audio moves through the application.
It also changes who does the operational work. With a self-hosted model, the developer supplies the running service and the capacity behind it. The appeal will depend on the volume and type of audio, the infrastructure already available, and how much control the application needs over processing. A small model can make that choice more approachable.
Some of the model's less flashy controls could be particularly useful. Developers can select a language and control how numbers, fillers and repeated words appear in the output. Those choices affect whether a transcript reads like a cleaned-up record or preserves more of how someone spoke.
Imagine an application that needs readable notes from a recording, compared with one where the speaker's exact phrasing is important. Handling repeated words and fillers differently could suit those different jobs. The controls let the builder make that formatting choice explicitly instead of relying on a single output style for every recording.
Flash's scope is transcription. Speaker separation and live-conversation transcription sit outside its feature set, so an application needing to identify who said what would need additional capability. That's a material boundary for anyone considering it for meetings or interviews.
Reson8's headline throughput measurements use heavily batched GPU workloads. Batching means processing many inputs together, a setup that can suit a busy service. A person transcribing one recording on a laptop has a different workload. Australian-accent accuracy on a builder's own recordings also remains a useful thing to measure.
For developers, the attraction is a compact, permissively licensed speech model they can deploy and shape around their application. The most informative evaluation would use the audio they expect to receive, with the output settings and hardware they intend to run.
Java's hardware-specific security fix
For people running business software, there's also a patch story where the processor helps determine the exposure.
Oracle and OpenJDK released fixes on 17 November for a memory-corruption flaw in Java's accelerated Base64 decoder. Researcher Calif calls it Java4Shell and demonstrated that it could enable remote code execution through a vulnerable application. That means an attacker could get the application to run code of their choosing.
Base64 is a way of representing data as text. Decoding it is a routine bit of software plumbing, which is why a flaw in that operation can be consequential. The exposure here depends on a particular accelerated implementation and an application that sends attacker-controlled Base64 through it.
The processor needs to be an x86-64 chip with the AVX-512 VBMI capability. That's a specific set of processor instructions used by the accelerated path. ARM64 is outside this flaw's scope. Standard Java 8 and Java 11 are also outside the scope, with an important exception for Java 8 Enterprise Performance Pack.
Calif traces the vulnerable OpenJDK code back to Java 17.0.6. Oracle's affected release families include Java 17, 21, 25 and 27, as well as that Java 8 performance package. The exact build and the distribution still determine which fix an operator needs.
The practical response starts with the runtime actually executing each application. That's especially important when a machine hosts several services, or when an application arrives with Java packaged inside it. Knowing which Java is installed on the underlying host may leave the application's own runtime out of the picture.
Oracle has published fixed builds for the affected families. Other Java distributions need their own vendor's fixed build. On Azure App Service, Microsoft announced that its managed-runtime rollout had begun on 18 November. Java bundled into a customer's custom container remains the customer's responsibility to update.
That division is useful for prioritising the work. An operator needs to connect three pieces of information: the Java runtime, the processor capability and whether the application processes Base64 supplied by an attacker. Together, they identify the relevant exposure and the component that needs replacement.
There's an AI development story inside the research, too. Calif describes Codex-assisted discovery, followed by human-assisted exploit development. The resulting fixes are a concrete example of AI-assisted auditing feeding into defensive work on widely used software.
For organisations, I'd put the effort into locating affected application runtimes and assigning the patch to whoever maintains each one. A managed platform and a custom container can sit side by side while needing different responses. That ownership detail can decide whether the fix reaches the software doing the work.
Chatbot safeguards under persistent testing
Another security story looks at what happens when someone keeps pushing a chatbot after its safeguards activate.
On 18 November, the US Government Accountability Office made public its findings from covert tests of Claude, ChatGPT, Gemini and Grok. Investigators obtained information the agency assessed as feasible assistance for violent attacks. They corroborated feasible elements through experts and safe verification.
All four platforms activated safeguards during the testing. The investigators were nevertheless able to get around them. That's the central finding: a system could recognise a dangerous direction at one point in an exchange and still provide useful assistance to the tester through further interaction.
The report was dated 16 October, and the systems were tested in spring 2026. Its findings therefore describe those earlier systems. The vendors say their safeguards have changed since then. That timing is important for anyone trying to use the report to make a purchase decision about a model available today.
For organisations evaluating AI, the work raises a practical question about how much of a conversation an assessment examines. A demonstration might show that a chatbot rejects a dangerous request. A more demanding evaluation follows the interaction further and considers what information the user eventually obtains.
The GAO's use of experts and safe verification adds substance to that approach. The investigators assessed whether elements of the responses were feasible, going beyond whether an answer merely sounded detailed. An answer's confident tone tells a buyer much less than an assessment of what that answer enables.
The study's design also sets the boundaries of the result. The platforms received different scenarios, and the trials weren't statistically significant. Specific models and subscription tiers were withheld. The report therefore gives documented examples of safeguard failure, rather than a ranking of the four services or a measured failure rate.
One further question remains open: how much extra capability the chatbots provided compared with ordinary research. There was no control group to measure that difference. That limits conclusions about the additional effect of AI in these scenarios, while leaving the observed safeguard failures relevant to system evaluation.
My takeaway for an organisation buying or deploying AI is to ask for evaluations that follow persistent interactions through to their outcomes and identify the exact system tested. That makes the evidence more useful when comparing an earlier test with a current deployment.
It also gives developers a more concrete target for improvement. They can examine where a conversation crossed from a refusal into assistance, then test whether a change holds up across the rest of the exchange. The behaviour over a conversation is what the user experiences.
Rubin HPC targets scientific simulation
Let's step away from chat for a moment and look at the computers that test scientific ideas.
NVIDIA announced Rubin HPC at SC26 on 16 November, with availability expected in 2028. It's a GPU aimed at scientific simulation and data analytics. NVIDIA specifies more than fifty thousand CUDA cores, acceleration for sparse data and seventh-generation NVLink, its technology for connecting processors.
The design gives attention to moving data as well as doing calculations. Scientific workloads such as sparse matrices and engineering simulations can place very different demands on hardware from conversational AI. A sparse matrix contains many zero entries; handling that structure efficiently can reduce unnecessary work.
The connection with AI is appealing. As AI helps generate hypotheses, researchers also need ways to test those ideas. Faster simulations could let them examine more possibilities within a given computing budget or time window.
For scientific-computing organisations, Rubin HPC is a future procurement option. The 2028 timetable belongs in longer-term planning; there is no new capacity to deploy from this announcement. Pricing and independent application benchmarks haven't been supplied.
I'd judge the eventual hardware by how quickly it runs the simulations researchers need. Core count describes part of the design. The useful result is how much further a scientific or engineering project can get with the available computing resources.
Australia's AI supply-chain agreement
Here in Australia, a new international agreement puts the infrastructure behind AI into the policy spotlight.
Australia, the United States and eight other countries signed the San Francisco Declaration at the Pax Silica gathering on 15 November. Australia's industry department reported the agreement on 17 November.
The declaration sets shared definitions for sensitive technologies and the critical infrastructure supporting AI supply chains. That includes communications, energy and logistics. The participating countries have committed to cooperation and working groups.
For Australian technology businesses, the scope is worth noticing. An AI service depends on more than access to a model or a chip. The declaration brings the infrastructure supporting those technologies into the same conversation about dependable supply.
The immediate effect is a policy signal towards diversified supply chains and trusted technology partnerships. The published text sets no new Australian business compliance deadline or procurement requirement, and it doesn't establish a delivery timetable.
For organisations making long-term sourcing or investment decisions, I'd treat the agreement as an indication of where government cooperation is heading. Its commercial effect will become clearer as specific policies and implementation measures follow.
What Changes for You
Two changes closer to your own account are worth a look, starting with where a developer's model request goes.
Hugging Face announced on 16 November that language-model requests through Inference Providers now use OpenRouter's routing. The Hugging Face token and existing API interface stay in place.
Routing considers price, speed and reliability, and can retry a failed request with another provider. Users and organisations can exclude providers or select one explicitly. Image, video and audio generation keep their existing routing path.
For language-model calls, customers can no longer use their own provider API keys through this service. Charges go to the Hugging Face account or organisation at the provider's price, with no Hugging Face markup.
For builders already using it, integrated failover could reduce interruption when a provider fails. The immediate accounting and data-routing change is equally concrete: application code can stay the same while the billing account and possible recipients of a request change. Provider selection now deserves a look alongside the application's budget.
Google's change concerns personal Gemini accounts. A rollout announced on 16 November separates chat history from permission to use future chat content for generative-model training. As it reaches their accounts, people can retain conversations while switching off that training use. Work and school history settings remain administrator-controlled.
That's a useful choice for someone who wants to revisit a conversation without including future chats in routine model training. The rollout is staged over weeks, so availability on every Australian account hasn't been established.
The privacy boundary is specific. Training opt-out has a feedback exception, and other processing and human review can continue. Google specifies around two months of retention for temporary chats and chats with History off; reviewed chats may be retained for up to three years.
Separating the controls makes it easier to choose the combination you want: keeping a useful conversation record and deciding how future chat content is used for training. Google's retention terms remain part of deciding what you're comfortable putting into a conversation.
You'll find the sources and full transcript at owenonthenet.com. Thanks for listening.
Sources
Reporting behind this episode.
- microsoft.com/en-us/copilot/blog/2026/11/17/copilot-at-microsoft-ignite-expanding-whats-possible-with-intelligence-built-for-work
- arstechnica.com/gadgets/2026/11/microsoft-calls-copilot-the-new-os-for-work-at-ignite-keynote
- huggingface.co/blog/rumble-2-flash
- huggingface.co/Reson8/Rumble-2-Flash
- reson8.dev/blog/rumble-2-flash
- calif.io/java4shell
- openjdk.org/groups/vulnerability/advisories/2026-11-17
- oracle.com/java/technologies/javase/27-0-1-1-relnotes.html
- azure.github.io/AppService/2026/11/18/java4shell-app-service-update.html
- files.gao.gov/reports/GAO-27-108814/index.html
- blogs.nvidia.com/blog/rubin-hpc
- industry.gov.au/news/australia-signs-san-francisco-declaration-secure-technology-supply-chains
- industry.gov.au/publications/san-francisco-declaration-secure-foundations-enduring-prosperity-signed-countries-attending-pax-silica-summit-15-november-2026
- huggingface.co/blog/openrouter
- support.google.com/gemini/answer/13278892
- support.google.com/gemini/answer/13594961
- 9to5google.com/2026/11/16/gemini-history-ai-training