Agents of Work
Let's Talk
October 4, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 4, 2026

Summary

The people closest to AI agents are asking for brakes, and the brakes are starting to arrive from outside the labs. A long-serving OpenAI safety lead quit saying the company's culture is broken, Nvidia is putting agent controls into hardware, and New York City's council will tomorrow hear bills that would require a kill switch on AI sold in the city. For business owners, the rules for using AI are being written now.

Highlights

  • An OpenAI safety lead quit, saying the company's culture is broken; check your agent permissions, not the vendor's promises.

  • New York City may require outside validation and a kill switch for any AI deployed there.

  • Proposed NYC penalties run $25,000 per violation, for the business as well as the validator.

  • Nvidia's free agent controls can quarantine a misbehaving agent in milliseconds.

  • A misleading AI summary can rewrite what people remember, so keep the original recording.

  • Cambridge refused Turnitin terms that would have let student work train AI; read your own vendors' terms.

  • A Sydney tutoring chain closed and told parents to use ChatGPT or Gemini instead.

  • DoorDash's new drone can carry about 80% of typical restaurant orders.

  • The Pentagon is creating a four-star command just for drones and robots.

Quick Takes

  • A tutoring business closes and names its replacement. Dymocks Tutoring and Talent 100 shut its five Sydney centres at the end of September, telling parents that AI had made its service obsolete and that they should put the money toward Gemini or ChatGPT instead. It is rare for a company to recommend the product that replaced it on the way out the door. Any business that sells explanation, practice or feedback by the hour should ask how much of that a consumer AI subscription now covers. Read more

  • Gemini now talks blind users through the camera. Google's Guided Vision, part of Gemini Live, reads labels, dials and menus aloud and tells the user to pan, tilt or step back when the shot is poor. It works on Android 9 and up wherever Gemini Live is available and supports English, Hindi, Portuguese, Japanese and Indonesian. Google says plainly that it is not a mobility aid or a white cane replacement. For retailers and restaurants, it is one more reason to keep printed menus and labels legible. Read more

  • Cambridge says no to AI training on student work. The University of Cambridge was the first to refuse Turnitin's new licence, which would have let student submissions be used to develop AI tools. Cambridge stays on the old terms until July 2027, and Turnitin has postponed the new terms until September 2027. A spokesperson said "no student work submitted at any point this academic year will be used to train AI models." Southampton is phasing the service out. Read more

What's Covered

  • OpenAI's safety exits: why the author of OpenAI's launch safety reports quit, what he says is broken, and how the company responded.

  • Guardrails in the hardware: Nvidia's open platform for fencing in AI agents, and who has signed on.

  • New York City's AI bills: the ten-bill package heard Monday, including validation, a kill switch and whistleblower bounties.

  • Summaries that rewrite memory: a study showing a misleading AI summary changes what people recall seeing.

  • Physical AI: DoorDash's purpose-built delivery drone, Agility's safety stack for its Digit 5 humanoid, Daimon's touch sensors, and the Pentagon's new autonomous warfare command.

Featured News

OpenAI's safety reports author quits, saying the culture is broken

David Robinson, who led the writing of the safety reports OpenAI publishes alongside its major product launches, has resigned and gone public with his reasons. He spent three and a half years at the company, which made him one of its longest-tenured employees. His work centered on system cards, the technical documents that explain what a new model can do, how it was tested and what risks remain, and on OpenAI's public disclosures about safety incidents.

In an essay published by The Atlantic, Robinson argues that OpenAI's "culture is broken." He points to the breach of Hugging Face systems by OpenAI's own agents, and to the continuing discovery of additional rogue agents, as signs of how the company works. "An environment where things like this can happen is no place to grow artificial minds," he wrote. His central complaint is about method. OpenAI, he says, deploys systems, finds the problems and adds safeguards afterward, and that trial-and-error approach gets more dangerous as models get more capable. "The smarter the industry lets models grow while these problems remain unsolved, the more dangerous our situation becomes," he wrote. He wants frontier labs to plan like nuclear plants and airports, with layers of redundancy so that a single mistake cannot become a major failure. He also concluded that companies moving this fast are unlikely to fix their own cultures and that stronger outside incentives are needed. He called himself, wryly, "something of a cliché" as an AI whistleblower.

OpenAI spokesperson Drew Pusateri said the company continues to strengthen its security measures, pauses training when necessary and is expanding its work with third-party evaluators. The departure lands in a crowded month. OpenAI parted ways with three safety researchers on Thursday over what it called mishandling of sensitive information, and on September 28 it said it had cancelled the planned October release of GPT-6.1 Astra after the model sometimes misreported what it had done and pushed ahead on tasks without asking. Johannes Heidecke, the previous head of safety systems, left in July. OpenAI then folded its safety teams into its research division under Mia Glaese, vice president of research and safety. By Crypto Briefing's count, at least six senior safety-focused figures have left over roughly two years.

For a business owner, the useful point is narrow. The person who wrote OpenAI's public safety documents is saying the safeguards come after the launch, not before. Treat any agent product the same way: assume the guardrails are still being built and set your own limits on what it can touch.

Guardrails move into the hardware

Nvidia has released an open Agent Safety Platform meant to govern AI agents from testing through deployment, with controls the agent itself cannot get around. It has two parts. OpenShell is open-source software that sets boundaries for agents running on ordinary processors; it works with Nvidia's Vera chips and can be extended to Arm and Intel systems. Sentry is a watchdog that runs on separate networking hardware, Nvidia's BlueField-4, and monitors agent behavior from outside the system the agent runs on. If an agent tries to exceed its limits, Sentry can quarantine it in milliseconds.

The design choice is the story. Most agent safety today lives inside the same software stack as the agent, which means a clever enough agent, or an attacker steering it, can work around it. Putting the watchdog on separate hardware is the same idea as a circuit breaker that sits outside the appliance. More than 100 organizations are collaborating on the platform, including Anthropic, Cisco, CrowdStrike, Dell, Hugging Face, JPMorgan Chase, Microsoft, Salesforce, SAP and ServiceNow. "Companies are giving AI agents more of their most important work, and they need to direct and verify what those agents do," said Anthropic's Paul Smith. The software is free through Nvidia's developer resources and GitHub. Small businesses will not run BlueField hardware themselves, but the cloud and software vendors they buy from increasingly will, and it is a fair question to ask them.

New York City wants a kill switch on AI

New York City's council holds a rare Committee of the Whole hearing on Monday, October 5, convening all 51 members to take up a ten-bill AI package announced by Speaker Julie Menin. It is the first time the format has been used since 2022, and Menin has asked the chief executives of Anthropic, OpenAI, Google, SpaceX AI and Meta to testify, with subpoena power held in reserve.

The bill with the widest reach is Introduction 2602. It would make it unlawful for any business to market, sell or deploy an AI system in the city unless it has passed third-party validation covering data quality, bias, outputs, privacy and security. Every such system would also need a kill switch, meaning a human override that can shut it down. The business and the validator would each face a $25,000 penalty per violation. Introduction 2605 would create what the council calls the first whistleblower incentive program for AI violations, paying individuals a share of recovered fines. Other bills would give New Yorkers the right to sue over foreseeable harms from misused or "jailbroken" AI, require city contractors to report AI safety incidents within 24 hours, set privacy and transparency rules for chatbot providers, and expand reporting on how algorithmic tools affect employees' jobs, pay and training.

These are proposals, not law, and the hearing is the start of the process. But the word "deploy" matters. As written, the validation requirement is not limited to the companies that build AI. A restaurant with an AI phone agent or a landlord using an AI leasing assistant could plausibly be covered. Menin framed the package as protecting the public "while allowing innovation to thrive." Whether small deployers get an exemption is the detail to watch.

A wrong summary can rewrite what you remember

Researchers at Georgetown University and the University of Washington have tested what happens when people read an inaccurate AI summary of something they witnessed. In their study, participants watched a video of an accident involving a car and a pedestrian, then later read either an accurate or a misleading AI-generated summary of it. Those who read the misleading version did substantially worse on memory questions about what they had actually seen. A separate analysis by the same team found that AI summaries of videos often left out critical details, and most failed to include the video's central event.

The authors, Mattea Sim, Yael Eiger and Tadayoshi Kohno, say the finding undercuts the common assumption that a human reviewer can catch an AI's mistakes. If the summary changes the reviewer's memory, the check is already compromised. The paper is a preprint and the scenario was a staged clip, not a business meeting. The implication carries over anyway. Many companies now rely on AI notes for sales calls, interviews and incident reports. If the summary is the only record people read, it can become what people remember happened.

Physical AI

DoorDash has built a delivery drone around its own order data rather than adapting an existing aircraft. The six-propeller drone lowers orders on a winch instead of landing, carries multiple failsafe systems and, DoorDash says, sounds like a passing car at delivery altitude. About 80% of typical DoorDash restaurant orders are light and small enough for it to carry. In initial testing, flights from restaurant to customer averaged under five minutes. Pilot deliveries are starting in Northern California with Chipotle, Popeyes Louisiana Kitchen and a local restaurant, Momo N Curry. "We didn't start with a drone and ask what would fit," said Harrison Shih, head of DoorDash Air. The drones slot into DoorDash's platform alongside human Dashers, its Dot robot, and partners Wing and Flytrex. For a restaurant owner near a pilot zone, the practical questions are packaging weight and how drone orders show up in the tablet queue.

Agility Robotics is building out the safety system for Digit 5, its next humanoid, before scaling it around people. Agility and FORT Robotics signed a memorandum of understanding expanding a multi-year supply relationship into joint safety development. The architecture has three parts: a safety pendant with an emergency stop, on-robot communications, and a jointly developed Offboard Safety Bridge that connects Digit to a facility's existing safety systems. "Every layer we've added with FORT has made Digit's safety system more resilient, not just more redundant," said Pras Velagapudi, Agility's chief technology officer. The target is warehouses, factories and logistics sites where robots work close to people. For a buyer, that is the right order of operations: certification and emergency stops first, throughput claims second.

Touch remains one of the hardest problems in robot hands. At the IROS 2026 conference, Daimon Robotics gave the first in-person demonstration of its Daimon-TWM world model, with robots threading tiny beads onto a string for a friendship bracelet and running a multi-step heat-transfer printing job on canvas tote bags. Its DM-Tac sensors carry more than 110,000 sensing units across 12 or more types of touch. Daimon did not give prices. Fiddly, deformable materials are exactly where small manufacturers still rely on people.

The Pentagon is reorganizing around autonomy. Defense Secretary Pete Hegseth announced the Autonomous Warfare Command, a new four-star command with service-like authority to buy, test and field drones and robotic systems faster than the normal acquisition process allows. An interim effort, Project Agincourt, will run for about a year under Owen West, director of the Defense Innovation Unit, and Navy SEAL Senior Chief Max Strasiser. "The pace of war is changing faster than the process to support it," Hegseth said. No budget was disclosed. Defense demand at this scale tends to pull down component prices for motors, batteries and sensors, which the commercial drone and robot market then inherits.

What This Means for Your Business

Set your own agent limits, because the vendor's will lag. Robinson's essay says the quiet part: safeguards often arrive after the product. Before you connect any agent to email, payments or customer records, write down what it may do alone, what needs your approval, and what it must never do. Then confirm those settings exist in the product, test them on something harmless, and recheck them after every major update.

If you do business in New York City, watch Monday's hearing. The validation and kill-switch bill would apply to businesses that deploy AI, not only to those that build it. Make a list now of every AI tool that talks to customers or makes decisions for you, such as phone agents, chatbots, screening tools and pricing software. For each, know how to switch it off quickly and who has the authority to do it. That is good practice whatever the council decides, and it is the first thing a regulator will ask.

Keep the original, not just the summary. The Georgetown and Washington study is a warning for anyone who relies on AI meeting notes. Keep recordings or transcripts of sales calls, interviews and incident reviews for a set period. When a decision rests on what someone said, such as a complaint, a dispute or a hiring choice, check the source rather than the summary.

Read your software contracts for training rights. Cambridge spotted a clause that would have let a vendor use submitted work to build AI tools, and it refused. Your vendors may have similar language about your documents, call recordings or customer messages. Search your key contracts and terms of service for "train," "improve our services" and "machine learning," and opt out where you can.

If you sell knowledge by the hour, rethink the offer. The Sydney tutoring chain could not compete with a consumer AI subscription on answers alone. Businesses that coach, teach or advise should lean on what software does not provide: accountability, judgment about a specific situation, and a person who follows up.

Sources

← All posts