mAIn Street #321: Nvidia adds enforceable limits to AI agents; Florida wants restrictions on OpenAI's model development; Google Gems will become reusable skills



Meta’s enterprise push, Nvidia’s agent controls, Florida’s court request and Gemini’s move from Gems to skills.
mAIn Street Daily Newsletter. AI news for people who actually have jobs to do.
Tuesday, September 29, 2026 · Edition 321

Good morning. Meta is building an enterprise AI business, Google is changing how saved instructions work, and Nvidia is putting controls around agents. Also inside: classroom experiments, accounting careers and new state rules taking effect this week.

TODAY’S THROUGHLINE

AI is moving deeper into business operations, bringing decisions about access, oversight and evidence to more workplaces.

TOP 5

Five stories worth your time.

Illustration of a student at a desk facing a robot that holds a lesson board.

Classroom AI is one of today’s five leading stories. Illustration: Julianna Brion for NPR. Read the reporting.

01

Nvidia / platform announcement · Sept. 28

Nvidia releases software to put enforceable limits around AI agents.

Nvidia’s Open Agent Safety Platform combines software boundaries with a hardware watchdog design. The aim is to enforce limits outside the AI model itself, giving organizations a way to monitor actions and stop agents that exceed their permissions.

02

Meta / company announcement · Sept. 28

Meta starts an enterprise AI business and hires CJ Desai to lead it.

Meta is creating a business unit to bring its AI models, agents and developer tools to companies. MongoDB CEO CJ Desai will lead it, reporting to Mark Zuckerberg. The move puts Meta more directly into the market for workplace AI.

03

The Next Web / court filing report · Sept. 28

Florida asks a court to restrict OpenAI’s model development.

Florida’s attorney general wants a temporary injunction requiring independent safety safeguards before OpenAI develops new models. The filing also seeks to keep minors in Florida off ChatGPT. These are requests to a state court, with no ruling reported.

04

TechCrunch / Google product notice · Sept. 28 report

Google will move Gemini’s Gems into reusable skills.

Google says it will automatically migrate Gems to skills starting November 17. Existing custom instructions will move with them, preserving work users have put into their assistants. The change was announced in Gemini and reported over the weekend.

05

NPR / KUOW · Sept. 28 reporting

Schools are testing classroom AI before they know how to judge it.

New NPR reporting finds districts trying AI for teaching, feedback and emotional learning while struggling to define success. A Stanford review of more than 800 papers found limited evidence to guide schools, leaving educators and families to assess experiments with uneven support.

AI Workflows | Tuesday: Context and harness management

A client handoff should carry the decisions that made the current plan.

On September 23, Google added a way to reference a Gemini Notebook inside Docs. A draft can draw on that notebook’s sources and show inline citations. For a client project, that opens up a useful handoff: the next person can see the source behind a deadline or promise without reconstructing months of conversation.

Imagine you manage a website launch and will be away for three days. The approved schedule says Friday. The client asked for Wednesday in a meeting, but your team has not accepted that change. A rushed summary could turn the request into a commitment. A good handoff keeps both facts, their dates, and the person who must decide.

Context: The signed scope, approved schedule, meeting notes, and dated decision log in the client’s authorized Drive folder.
 
Harness: Rules for adding new sources, labeling approved versus pending changes, citing commitments, limiting access, and stopping before any client message or schedule change.
 
Result: A one-page handoff with the current commitment, open requests, owners, next actions, and links to the underlying records.

Make one notebook for the client and give it only authorized project sources. Workspace Studio can add new Drive files as notebook sources when they enter the client folder; Drive-backed sources can sync later edits. Record the approval rule in the notebook’s instructions and in the handoff assignment. When ownership changes, @mention the notebook in Docs and have Gemini produce the cited handoff. This Docs feature requires an eligible plan, including Workspace Business Standard or Plus, with the relevant settings enabled.

Assignment to copy after selecting the notebook: “Draft a one-page internal handoff. State the latest approved deliverable and date, with a citation for each commitment. List newer client requests separately until an authorized person approves them. Name the owner, next action, and any missing decision. Check the signed scope and dated decision log against the meeting notes; flag conflicts you cannot resolve from sources. Do not message the client or edit the project schedule. I will check the citations and approve sharing.”

After you check the citations and approve sharing, give the handoff to the incoming owner. Ask them to find the approved deadline and the unresolved request. If either answer lacks a source, fix the notebook or intake rule before the next handoff. Managers can use that same two-question check across client teams.

AI TOOL SPOTLIGHT

Databox brings business context into its AI analysis.

Databox’s AI Analyst works with connected business data to answer questions, build metrics and produce reports. Its connectors can add the context behind your numbers, while saved metric definitions help keep recurring analysis consistent.

Try one question you already know how to check: which campaign brought in the most qualified leads last month? Connect the approved records, define “qualified,” and compare the answer with your existing report before scheduling it. The value depends on the data, definitions and access you provide.

Databox / features and trial details

THE DAILY ROUNDUP

Headlines

Business and working life

Meta / company announcement · Sept. 28

Meta starts an enterprise AI business and hires CJ Desai to lead it.

Meta Enterprise Platform will turn products including Muse, its business agent and developer tools into services for companies. Desai joins from MongoDB; the announcement establishes a new business unit, with the broader product offering still taking shape.

TechCrunch / company profile · Sept. 28

MAVI recruits accountants who can direct AI and check its work.

The talent marketplace connects U.S. companies with finance professionals worldwide for ongoing roles. Its hiring pitch combines accounting knowledge with the ability to use AI, spot mistakes and exercise judgment as routine work becomes easier to automate.

Bank of America / product announcement · Sept. 28

Bank of America adds AI-assisted payment analysis for business clients.

Payments Insights joins the CashPro analytics suite for corporate and commercial clients’ U.S. accounts. It uses payment patterns and peer comparisons to help treasury teams examine payment efficiency, cross-border transactions and working capital.

TechCrunch / funding report · Sept. 28

Outmarket raises $34.5 million to automate insurance brokers’ paperwork.

The company’s software reads insurance documents and helps brokers compare coverage and prepare applications. Its new funding comes four months after its previous round, backing a business built around reducing the document work behind commercial insurance.

Lofty / announcement via GlobeNewswire · Sept. 28

Lofty lets real-estate firms connect their own AI agents to its platform.

The new connection can give approved assistants access to brokerage records and authorized actions. A firm could use it to find promising leads that have gone without follow-up, using its existing business data and instructions.

Tools and control

Nvidia / platform announcement · Sept. 28

Nvidia releases software to put enforceable limits around AI agents.

OpenShell can record actions and control an agent’s access to systems. Nvidia also introduced Sentry, a separate hardware watchdog design intended to isolate agents that cross their boundaries; the watchdog is a reference design, while OpenShell is broadly available.

TechCrunch / Google product notice · Sept. 28 report

Google will move Gemini’s Gems into reusable skills.

Gems remain usable before the move. After migration, users will select skills with a forward slash in a task thread, changing how they call up saved instructions.

Lenovo / product announcement · Sept. 28

Lenovo’s first Googlebook puts Gemini tools into a lightweight laptop.

The Googlebook 15 is scheduled to arrive October 4, starting at $1,099.99. Its features include contextual actions through the pointer and dictation that removes filler words and adds formatting; availability varies by market and language.

TechCrunch / funding report · Sept. 28

Modulate raises $25 million to expand AI call analysis.

Its voice systems can examine calls for scams, synthetic voices and possible policy violations. The company is extending technology developed for online voice moderation into business calls, where a transcript alone can miss important context.

TechCrunch / company profile · Sept. 28

DetectifAI is building phone-based defenses against cloned voices.

The startup wants handset makers to embed small detection models into phones so suspicious calls can be checked on the device. Its proposed protection targets impersonation scams; it is a technology licensing effort rather than a newly available consumer app.

CSO / Microsoft security findings · Sept. 28 report; Sept. 25 disclosure

Microsoft details a rapid cloud attack linked to AI-driven extortion.

Stolen application credentials let attackers delete Azure resources in minutes. Microsoft linked the activity to Jadepuffer; the evidence shows coordinated automation, but does not establish that AI directed every action. Deleting an exposed password from a public post also left it visible in the edit history.

Rules and public decisions

The Next Web / court filing report · Sept. 28

Florida asks a court to restrict OpenAI’s model development.

The motion belongs to Florida’s existing lawsuit. OpenAI says it favors industry-wide rules instead of restrictions aimed at one company; the requested injunction is not a court order.

Reuters / Yahoo Finance · Sept. 28

Trump plans Tuesday talks with tech leaders about AI oversight.

House Speaker Mike Johnson says the White House meeting will weigh innovation against oversight. His comments oppose a development moratorium while calling for more transparency; the meeting’s outcome was still pending at this edition’s reporting cutoff.

CT Mirror / effective-date guide · Sept. 28

Connecticut begins another phase of its AI rules on October 1.

New AI subscription disclosures and frontier-model whistleblower protections take effect Thursday. Employment and companion-chatbot provisions have separate, later deadlines, making the effective date of each requirement an important part of the rollout.

Reuters / The Straits Times · Sept. 28

OpenAI and Anthropic decline Australia’s October 1 AI hearing.

Both companies cite short notice for the Senate committee appearance sought after the Medicare incident. OpenAI says its chief strategy officer will attend a separate parliamentary AI hearing on October 6, keeping the two proceedings distinct.

Axios / AllSides · Sept. 28

AI companies and unions join a data-center coalition.

The American Infrastructure Alliance seeks local standards to avert construction moratoriums.

Schools, communities and culture

NPR / KUOW · Sept. 28 reporting

Schools are testing classroom AI before they know how to judge it.

NPR’s reporting contrasts districts expanding experiments with others holding back. In Dayton, an emotional-learning chatbot asks students for advice and grades their replies; even the pilot’s evaluator says measuring success remains difficult.

Lenfest Institute and OpenAI / joint announcement · Sept. 28

Lenfest expands its program placing AI engineers inside local newsrooms.

OpenAI is committing $5 million plus up to $5 million in software credits and engineering support. The next phase will add news organizations to a program whose fellows have built archive-search, translation and reporting tools alongside newsroom staff.

University of Pittsburgh / grant announcement · Sept. 28

Pittsburgh will train coal workers and electricians for new technical jobs.

A $1.62 million Appalachian Regional Commission award supports laser-processing training for displaced coal workers and fiber-optic sensing training for electricians. The latter includes monitoring AI data centers and power systems; the project aims to train at least 320 people over three years.

NHL / partnership announcement · Sept. 28

The NHL brings Adobe’s AI tools into its fan-content operations.

The league and its 32 clubs will use Adobe tools to produce and personalize content across audiences and languages. Fan-facing creative experiences are also planned, but those are future offerings rather than tools supporters can use immediately.

Health and research

Investing.com / company presentation report · Sept. 28

Roche is building autonomous AI laboratories for drug research.

The company says the labs are part of its effort to speed research and development. It also reports that AI or computational work informed 40% of pipeline decisions over a recent three-quarter period; that measures involvement in decisions, not successful new treatments.

IDIBELL / research release via Medical Xpress · Sept. 28

Researchers use AI to distinguish young and aged mouse blood stem cells.

ChromAgeNet examines the organization of DNA in microscope images and could help screen potential interventions. The study’s treatment experiments showed changes in cell appearance, without demonstrating restored cell function or a treatment for aging in people.

Northeastern University / researcher interview · Sept. 28

Northeastern researchers call for stronger testing of AI companions.

Their peer-reviewed analysis describes risks including dependency and reinforcement of harmful beliefs. In a new university interview, the researchers argue for more rigorous evaluation; the paper does not establish how often these harms occur across all chatbot users.

Purdue University / research report · Sept. 28

A Purdue study finds that preparing science for AI also changes scientists’ work.

Interviews with 26 researchers and program managers show how automation projects break expertise into smaller tasks and standardize methods and data. The analysis examines changes in scientific practice without assuming that fully autonomous science will succeed.

Chief Healthcare Executive / interview · Sept. 28 publication

Hospital accreditors want ongoing checks on AI in patient care.

In a newly published interview, Joint Commission leaders describe a role for AI in preventing handoff errors. They emphasize existing governance principles, including patient transparency, incident reporting and repeated evaluation as hospitals put the technology to work.

mAIn Street gives nontechnical readers useful AI news they can put to work.

mAIn Street #321 · Tuesday, September 29, 2026

Reporting checked through September 28, 2026.

​Read All mAIn Street Back Issues Here​


mAIn Street

Check out the resources I offer below and sign up for my new newsletter!

Read more from mAIn Street
Microsoft presents Copilot’s Home, Code and Autopilot interfaces side by side.

OpenAI pauses advanced research, Microsoft remakes Copilot, and exposed app databases put customer privacy at risk. Plus work, schools and early cancer research. Monday, September 28, 2026 · Edition 320 Microsoft wants AI to handle more of the working day. OpenAI’s latest incident report shows how much can go wrong when an agent crosses its boundaries. This Monday’s briefing also looks at practical tools for accounting and customer service, questions facing schools, and research that could...

Four people wearing styles from Meta’s newly announced AI glasses range.

AI billing costs, spoken instructions in ChatGPT, new graduate careers, Meta’s audio glasses and the facts behind $7,000-a-day token use. Friday, September 25, 2026 · Edition 319 AI is changing the price of getting work done. Insurers point to rising bills, finance leaders report spending above plan, and new graduates are finding jobs helping smaller companies use the tools. Also inside: spoken assignments in ChatGPT, Meta’s audio glasses and the source behind those $7,000-a-day token...

YouTube creation tools illustration with editing and Studio feedback screens.

Amazon’s seller automations, GPT-6 Sol and Luna, Gemini app connections, YouTube’s creator tools and a federal crash investigation. Thursday, September 24, 2026 · Edition 318 Amazon is giving sellers recurring automations. Google is bringing more familiar apps into Gemini, and OpenAI has released cheaper models for metered work. This edition follows those changes into the workplace, alongside factory workers training robots, new fraud defenses and a federal investigation into fatal crashes...