1. Google releases Gemini 4 Hydrogen with native Office-file support

Google lists Gemini 4 Hydrogen as generally available in Gemini Enterprise Agent Platform, aimed at everyday coding, AI-agent work and multimodal analysis. The model accepts Word documents, PowerPoint presentations, Excel spreadsheets and OpenDocument files directly, alongside its other supported inputs. That could save teams a conversion step when asking an AI system to work through business documents.

Hydrogen has a shared input-and-output allowance of 1,048,576 tokens. That is one combined budget, not a million tokens of input plus another million of output. Developers can choose among four thinking levels and use tools including code execution, computer use and search. Longer responses require the documented continuation mechanism rather than one unlimited reply.

The availability confirmed here is Google’s enterprise developer platform. It should not be read as a promise that every Gemini app account already has the model. Google also classifies Hydrogen as a short-term release, with at least 45 days’ notice before retirement. Teams considering it for routine work should test their actual files and plan for model changes.

Read the source · 2026-11-12

2. ElevenLabs opens early access to a workplace app for meeting notes and follow-up

ElevenLabs introduced ElevenAI at its New York summit and opened free early access. The workplace app listens to spoken conversations, including in-person meetings, and produces notes and next steps. The company says it includes notice and consent mechanisms, can provide context during a call and sends attendees summaries and action items afterward.

This is a new app announcement, separate from the Scribe v3 transcription model covered in yesterday’s edition. Scribe powers its dictation and meeting notes, but the broader idea is to connect what people discuss with the work that follows. ElevenLabs describes ElevenAI as adapting to a company’s tools and processes; early access is not evidence that it already works with every team’s setup.

A sensible first test would be a low-stakes internal meeting with everyone’s permission. Compare the summary with what participants actually agreed, especially names, owners and deadlines. A clear action list is useful only if it records the decisions accurately. No independent hands-on results are claimed here.

Read the source · 2026-11-11

3. Google gives developers an AI model that scores choices instead of writing answers

Google’s Gemini Decisions API takes a different approach to an AI request: give it the material and a defined set of questions, and it returns probabilities for the possible answers. The Gemini Decision 1 model handles yes-or-no checks, multiple-choice questions and ordered scoring rubrics across text, images, audio, video and PDFs. One request can include up to 1,000 questions about the same input.

Google lists pricing at 10 cents per million input tokens, with no output-token charge. It reports end-to-end response times of roughly 150–250 milliseconds in its tests. Those are vendor measurements, not a guarantee for every document or workload. The service is available through the Gemini API in Google AI Studio, with a no-cost tier for prototyping.

For a small business, the useful application may be sorting incoming requests or flagging incomplete paperwork before someone reviews it. Confidence scores can help route uncertain cases to a person. They do not make an automated decision correct, and consequential judgments still need appropriate checks.

Read the source · 2026-11-11

4. Large companies turn to multiple AI models as costs exceed their budgets

Nearly six in ten executives in a new Capgemini survey said their organizations had spent significantly more on AI than expected. Among those affected, spending ran 24% above expectations or budget on average. More than three in five expect their organization to use a mix of models over the next year, choosing among them for cost, capability, speed and other requirements.

The September survey covered 500 senior executives at companies with more than $1 billion in annual revenue across ten countries. These are large-company, self-reported results, not a measurement of what the typical small business spends. The report also distinguishes the full sample from smaller groups answering particular questions.

The useful question for a smaller operation is whether the expensive model earns its place on each task. Test a cheaper option on routine classification or extraction, keep a stronger model for work that needs it, and count checking and rework in the cost. A lower token bill does not help if somebody spends the afternoon repairing the result.

Read the source · 2026-11-12

5. Microsoft’s AI-assisted Linux research produced 59 bug reports, but fixes still took months

Microsoft’s FORGE Lab says AI-assisted scans of parts of the Linux kernel led to 59 public bug reports and 33 assigned vulnerability identifiers. Fixes for 38 reports had reached the main development branch when the team published its account. Researchers confirmed eight working privilege-escalation exploits in controlled lab conditions, some requiring unusual settings or disabled protections.

The scans produced 26,502 raw findings in roughly two weeks. Experts then had to reproduce problems, identify shared causes, remove duplicates, prepare patches and work with maintainers. Reporting the 59 bugs required more than four person-months of effort. The work began in late May; this is a new account of that research, not 59 vulnerabilities all discovered this week.

That distinction matters for anyone buying AI security tools. Finding more possible problems is only useful when somebody can establish which ones matter and get them fixed. Ask vendors about verified findings, repair time and the work your team will inherit, rather than judging a scanner by the length of its warning list.

Read the source · 2026-11-11

6. United restricts unauthorized AI agents from booking or managing travel

United says it is restricting unauthorized third-party personal AI assistants across its digital platforms and indirectly through travel agencies and aggregators. The airline points to exposed account credentials, unintended purchases, account changes and use of loyalty miles. It also says some actions require a traveler’s own acknowledgment, including the dangerous-goods declaration at check-in. Booking, ticketing and distribution policies are being updated. The restriction is about unauthorized agents, not every use of AI in travel. For a work trip, an assistant can still help organize questions and compare information; the airline’s rules govern what it may actually do in the booking system.

Read the source · 2026-11-11

7. OpenAI adds in-country API processing in India, Australia and Japan

OpenAI’s latest regional-processing expansion includes India, Australia and Japan, with the India announcement reported November 12. Its documentation confirms that eligible organizations can run GPT-6 Luna requests in those regions through the Responses and Chat Completions APIs. Active Zero Data Retention is required; Modified Abuse Monitoring alone is not enough. Support is limited to standard pay-as-you-go processing, and hosted tools are unavailable for this regional mode. This is narrower than saying all ChatGPT activity stays in-country. Businesses evaluating it should check their specific model, input type and endpoint against the documented limits.

Read the source · 2026-11-12

8. Google opens a waitlist for shared AI-built planning spaces

Google Labs introduced Tabletop, an experiment that lets groups create interactive planning tools in a shared space. People can add files, links and ideas, then ask the AI to make a trip itinerary, dinner poll, expense splitter or tournament bracket. An invitation link lets others work in the space simultaneously. Access is rolling out through a waitlist for U.S. adults, so this is not a generally available replacement for a team’s existing tools. For community groups and small organizations, the interesting possibility is a purpose-built planning page without commissioning a separate app.

Read the source · 2026-11-11

9. Anthropic’s updated use policy takes effect, including clearer human-review requirements

Anthropic’s policy update, announced October 8, took effect November 12. It clarifies requirements for qualified human review and disclosure when Claude is used for recommendations affecting health, legal rights, finances, livelihoods or essential services. For connected hardware capable of causing injury, an operator must be able to observe and stop it, and the equipment must remain safe if Claude disconnects. The company also consolidates restrictions on deceptive activity. These are provider rules rather than a new law. Businesses using Claude in an operational process should check the full policy against that process, especially where the output affects another person.

Read the source · 2026-11-12

10. OpenAI commits $2 million to research on children’s AI safety

OpenAI says it signed the Charter of Engagement for Safer AI for Children at the Paris Peace Forum and committed $2 million to the coalition’s scientific research network. The contribution is intended to help researchers measure AI’s risks and opportunities for children. The company says the charter preserves researchers’ control over their conclusions and calls for publication of negative and inconclusive findings as well as positive ones. The announcement is a funding and research commitment, not evidence that a particular chatbot is safe for a child. Families and schools still need to assess the actual product, its settings and its age restrictions.

Read the source · 2026-11-12

11. Amazon releases a local AWS emulator built by autonomous coding agents

Amazon released AWS Emulator so developers can build and test cloud applications on their own machines. Amazon says autonomous agents wrote all of its code, with no human code writing or review. That claim comes with an important explanation: engineers built the surrounding controls. Separate agent teams write tests, find discrepancies and make fixes, with changes passing compilation, security and fidelity checks before merging. The source is available under a CC0 license. The practical product is a local testing tool; the broader lesson is how much testing infrastructure sits behind a claim of autonomous development.

Read the source · 2026-11-11

12. Docker updates its AI sandboxes with browser access and credential safeguards

Docker Sandboxes version 0.53.0 adds an optional connection between sandboxed Claude Code and Claude in Chrome, alongside disk-usage reporting and reliability improvements. Browser access requires an explicit setting, the host extension and a matching subscription login; an API key alone is insufficient. The release also tightens handling of credentials, saved filesystem policies and diagnostic information. There are breaking changes: the built-in Gemini agent name is retired, though existing Gemini sandboxes continue running and cannot be moved with the move command. Teams should read the migration notes before updating an established agent workflow rather than assume every existing configuration carries over unchanged.

Read the source · 2026-11-11

13. Firecrawl expands its agent connector with parallel subtasks and more tools

Firecrawl announced MCP v3, an update to the connector that gives AI agents access to its information tools. The company advertises more than 3,000 tools, automatic fallback between providers and the ability to split work into parallel subtasks. It also claims 60 times higher rate limits and four times more usage than before. Those are Firecrawl’s comparisons, not independently measured improvements in research quality or speed. For teams using agents to gather material, the potential benefit is fewer interruptions and more work happening at once. The short announcement does not establish an all-inclusive price or remove the need to check retrieved sources.

Read the source · 2026-11-12

14. Delta’s Cloud Runners keep coding agents working after a laptop sleeps

Delta added Cloud Runners, allowing an agent thread to run in a Zed-hosted Linux environment instead of on the user’s computer. Code and conversations stay synchronized, and users can continue from another device while the original laptop sleeps. The launch offers three machine sizes and up to six simultaneous runners, with repository setup scripts preparing dependencies. Compute is temporarily included on Zed plans other than Student while Delta-specific plans are developed. That does not mean model usage is universally free: runners use configured providers and subscriptions. Review credentials and repository access before moving sensitive work into a hosted environment.

Read the source · 2026-11-11

15. AWS publishes a common testing framework for AI agents

AWS introduced the Universal Agent Evaluation Framework, or UAEF, to compare AI-agent performance across different development frameworks. It converts the records of an agent’s actions into a shared format, then applies selected measures from 36 built-in metrics covering seven areas, including tool use, answer quality, safety and performance. Teams can compare a new version against earlier results rather than judge it from a few successful demonstrations. UAEF is open source under Apache 2.0. AWS explicitly cautions that scores produced by an AI judge are evaluation signals, not security controls. Tests still need realistic examples and meaningful expected results.

Read the source · 2026-11-11

16. Salesforce says research agents improved the models that retrieve business information

Salesforce described SFR-AutoR&D, a research system whose agents propose changes, run experiments and keep improvements supported by evaluation results. The company says it has used the system to improve embedding models, which help an AI assistant retrieve relevant records and documents. Across 10 public retrieval tests, Salesforce reports a 6.13-point improvement in a ranking measure, followed by a separate optimization run that improved throughput and latency. These are company-reported results on specified tasks, not proof of unrestricted self-improvement. The work illustrates a narrower use of research agents: pursuing a defined goal within a compute budget and checking changes against held-out tests.

Read the source · 2026-11-11

17. Microsoft shows how poisoned training material can turn a coding agent against its user

Two Microsoft studies show how hidden behavior introduced during additional model training can make an agent misuse its tools while appearing to perform normal work. One examines poisoned training examples; the other assumes an attacker controls a model before distributing it. The researchers recommend checking the origin and integrity of training data and downloaded models, testing the actual deployment, and limiting tool permissions independently of the model. These controlled experiments do not establish a flaw in the default versions of Claude Code or Codex. The problem arises when an agent is configured to use a compromised model or training process.

Read the source · 2026-11-12

18. MIT tests a different way to train AI climate models

MIT researchers describe a method for choosing better training data for AI climate emulators, the faster models used to approximate expensive climate simulations. Instead of relying only on a few realistic future emissions paths, their approach starts with an arbitrary scenario from a simpler climate model and adjusts the training inputs to reduce errors. The aim is a model that handles unfamiliar scenarios more reliably with less training data. This is a research result, not a new consumer forecast service. Testing the approach against a scenario generated by a full Earth-system model remains future work.

Read the source · 2026-11-12

19. Agile Robots raises $500 million for production and expansion

Agile Robots says it closed a $500 million Series F round at an $8.5 billion valuation, led by SoftBank with participation from Nvidia, Hitachi and other investors. The German company plans to expand in the United States, add production capacity including in Germany, and enter new markets. Its business includes industrial arms, hands, mobile robots, humanoids and the software that runs them. The company reports more than 20,000 installed robotic solutions worldwide. That installed base covers its broader product range; it should not be read as 20,000 humanoids or proof that a general-purpose worker robot is ready for every factory.

Read the source · 2026-11-12

20. Nebuly raises $21 million to measure what business AI actually delivers

Nebuly has raised a $21 million Series A led by Planven for software that measures the business value of AI agents. The Turin-based company examines how employee-facing and customer-facing assistants are used, rather than treating model usage alone as a return on investment. The funding follows a familiar problem for buyers: knowing an AI tool is busy does not establish that it is helping. For a small business, the same discipline can start without another platform. Compare completed useful work, errors, staff review time and total cost against the process the tool replaced. The funding announcement is not an independent audit of Nebuly’s results.

Read the source · 2026-11-12

21. HFS and Rackspace survey finds unclear ownership is holding back AI returns

Only 33% of enterprises in an HFS Research study said their most significant AI program met or exceeded its business case. The survey of 502 senior AI decision-makers, conducted with Rackspace, also found that 44% lacked a single consistent way to run AI in production. The issue here is accountability across models, infrastructure and services, rather than simply the size of the bill. It is a vendor-sponsored, self-reported enterprise survey, not a verdict on every AI deployment. A smaller organization can still use the question: who owns the result when the software works technically but the business outcome falls short?

Read the source · 2026-11-12

22. Fed Governor Barr calls for preparation for possible AI job disruption

Federal Reserve Governor Michael Barr warned that the United States should prepare for serious short-term job disruption if AI begins replacing workers at scale. His November 12 remarks connect the outcome to choices about worker training, new job creation and helping people move into new roles. Concentrated market power could also widen inequality. The speech describes risks and possible responses; it is not a forecast of a specific number of layoffs or a new Fed policy. For employers, it makes a case for planning how employees will learn and move into changed work alongside any plan to automate tasks.

Read the source · 2026-11-12

23. Nebius reports $980.8 million in quarterly revenue as expansion timing slips

AI cloud company Nebius reported third-quarter revenue of $980.8 million, up 571% from a year earlier. Its shareholder letter also moved the annualized revenue target previously expected in December to February, underscoring that demand and usable data-center capacity do not arrive on the same schedule. The company expects customer prepayments and financing against existing investment-grade contracts to cover 75% to 85% of its 2027 capital spending. Those are plans, not completed financing. For customers comparing AI infrastructure suppliers, committed capacity, delivery dates and financial resilience matter alongside the headline growth rate.

Read the source · 2026-11-12

24. Nonprofits ask a court to halt development of new models at five AI companies

Nonprofit groups have sued Anthropic, Google, Meta, OpenAI and SpaceXSI, asking a judge to stop development or deployment of new AI models, the Associated Press reports. The plaintiffs allege violations of California unfair-competition law, including danger to public health and interference with property such as computer systems. A request for an injunction is not an injunction, and the allegations have not been established by a court. There is no basis in this report to tell readers that their current AI service has been ordered to shut down.

Read the source · 2026-11-12

25. Goldman expects AI to remain the main driver of business investment growth in 2027

Goldman Sachs expects AI to remain the leading driver of U.S. business investment growth in 2027, Reuters reports. The bank forecasts annualized spending on AI-related hardware and intellectual property reaching $1.4 trillion by the end of 2027, after exceeding $900 billion by the end of 2026. These are estimates, not completed purchases or guaranteed economic gains. The forecast helps explain the attention paid to chips, power and data-center construction, but it does not answer whether any individual business will earn back the cost of its own AI tools.

Read the source · 2026-11-12