Agents of Work
Let's Talk
November 7, 2026 · Agents of Work

Agents of Work AI Daily Briefing — November 7, 2026

Summary

The biggest platforms have decided that if an AI agent is going to shop, book or plan for their customers, it will be their own agent, not a general-purpose one. Amazon, Airbnb and Meta all made that bet this week, while AI agents kept finding old security flaws faster than humans can patch them and the money behind the frontier labs got a little more cautious.

Highlights

  • Amazon now lets invited US shoppers text Alexa to search, compare and buy, with every purchase approved first.

  • Airbnb will ship its own travel agent in the first half of 2027 rather than rely on outside agents.

  • Expect each big store to push its own helper; outside agents have placed only hundreds of DoorDash orders.

  • Google DeepMind's agents found two new weaknesses in old 1024-bit RSA keys; inventory any legacy keys.

  • An AI-agent probe found a decade-old OpenSSH login bypass on Alpine Linux; patch to 10.6p2.

  • More than 800 copies of the Artex AI attack tool are exposed online; weak passwords are still the way in.

  • OpenAI's new round is reportedly $11 billion at $1.25 trillion, well short of its September target.

  • Meta's Muse now builds custom trackers and planners from a sentence and talks by voice.

  • Tesla will wind down the Cybertruck in mid-2027 to make room for Cybercab robotaxis.

  • Axon is putting $1.2 billion more into Skydio drones, which now arrive before officers on most calls.

Quick Takes

  • Baidu releases a bare foundation model for builders. Baidu's ERNIE-5.5-Flash-Base is a 280-billion-parameter mixture-of-experts model with 14 billion active per input, pre-trained on 33.5 trillion tokens and shipped deliberately without fine-tuning or post-training. The weights are on Hugging Face in FP8 and BF16, and Baidu says a post-trained version will follow. It is another sign that capable open models from China are arriving on a steady cadence, which keeps pressure on US model prices. Read more

  • Cursor will build, host and publish a small app for you. Cursor's new Live Apps feature lets a user describe a tool to a cloud agent, which scaffolds the web app, spins up a Postgres database, puts sign-in behind Cursor SSO and provisions hosting. Security checks run first, and "Nothing goes live until the checks pass and you explicitly publish the app." Team admins can see every app and secret, add rules to the review and unpublish anything. It is on all paid plans with cloud agents enabled. For a small team that needs an internal tracker or intake form, the cost of trying has dropped to an afternoon. Read more

What's Covered in Featured News

  • Every platform wants its own agent: Amazon's text-to-shop Alexa, Airbnb's 2027 travel agent, DoorDash's numbers on outside agents, and Cloudflare's plan for telling agents apart.

  • AI agents find old flaws fast: DeepMind's RSA-1024 results, an OpenSSH bypass found by agent probing, and hundreds of exposed Artex attack tools.

  • The money gets cautious: OpenAI's smaller round, lower expectations for Anthropic's IPO, Dropbox's $350 million Tigris deal, and Google taking over Crusoe's Wyoming site.

  • Meta's Muse grows up: Spaces that build trackers from a sentence, voice conversations, and 10 million users.

  • Claude trades in a video game: what Anthropic learned about long-running agent plans and catching a saboteur on the team.

  • Physical AI: Tesla ends the Cybertruck for Cybercab, Axon's $1.2 billion Skydio stake, Wing's US factory, Xiaomi's 8-hour humanoid shift, a $60-a-week robot house cleaning, and Nvidia's VFlash.

Featured News

Every platform wants its own agent, and it starts with shopping

Amazon on Friday began letting a small group of US customers shop by texting Alexa. Invited shoppers get a text asking them to enable the feature; once on, they can send plain-language requests to search for items, add them to a cart and buy, as well as track prices, compare products and check order status. Users can text a photo of an item to find similar listings. Amazon says everyone else can join a waitlist and that the feature will roll out more broadly in the coming weeks, in time for the holiday season.

The timing is the story. In September, Amazon blocked Meta's Muse agent from querying its site. Now it is offering its own agent inside its own store. Airbnb made the same move a day earlier, telling shareholders that its AI travel agent will launch in the first half of 2027. CEO Brian Chesky said general-purpose agents are fine for "simple routine tasks like making a restaurant reservation," but that "Travel is aspirational," and 80% of travelers enjoy the research. Airbnb's agent will assemble its answers from "a library of hundreds of travel-specific interface components," and the shareholder letter says it will handle reservation changes, payment issues, refunds, rebooking and 24-hour safety support. Airbnb can afford the bet: third-quarter revenue rose 19% to $4.87 billion, its fastest growth in more than three years. DoorDash, meanwhile, said last month that third-party agents had placed only hundreds of orders in total on its platform, and it is focusing on its own text-based agent.

Cloudflare published the infrastructure side of this fight on Thursday. Its post argues that when a hosted agent such as Muse or OpenAI's dots acts for a user, the site can no longer recognize the person or hold them accountable, and it lists 14 problems, from account sharing to friendly fraud, in which "the Site bears the primary cost." Cloudflare is backing a standard called Web Bot Auth for agent identity, is taking a Trusted Agent Delegation coalition to the IETF, and now lets customers see personal-assistant traffic and set per-agent access rules.

For a small retailer or service business, the agents customers meet first will belong to the biggest platforms, and those agents favor their own catalogs. The tools to see and manage agent traffic on your own site are just arriving.

AI agents are finding old security flaws faster than anyone expected

Google DeepMind announced on Friday that AI agents, steered by expert cryptanalysts, found two new weaknesses in legacy 1024-bit RSA encryption in a matter of weeks. The work was led by Dj Dvijotham and Abhradeep Thakurta with Nadia Heninger of UC San Diego and Emmanuel Thomé of Inria. On signatures, the agents showed that once an attacker controls roughly half of a signed block, padding adds almost no protection, and reproduced a known forgery on an already-solved benchmark key with 3.5 times less CPU time. For the harder case where an attacker controls only a 32-byte fingerprint, they found an algorithm DeepMind estimates is about 1,600 times better than the previous best attack, a theoretical figure rather than a measured break. A second result cuts the cost of factoring each additional 1024-bit key in a batch by about 30 times. Modern 2048-bit and larger keys are not affected. The risk is in old keys that linger in firmware signing, VPNs and internal certificate chains.

The same week, researcher Ozgun Kultekin disclosed an OpenSSH flaw that he found by running about 17,600 authentication tests with AI agents. On Linux systems that use the musl C library, including Alpine, OpenWrt, Buildroot and Yocto, an attacker can log into an account whose password is disabled simply by submitting a password longer than 256 bytes. The bug has existed since 2016. Alpine's official AWS and Azure images ship with an affected default account that can reach root. Debian, Ubuntu, Fedora and PAM-based systems are not vulnerable. The fix is OpenSSH 10.6p2; until then, disable password authentication.

Attackers have their own tooling. Censys found more than 800 copies of Artex, an open-source Chinese AI penetration-testing system, exposed to the internet without a password. Artex has been linked to attacks on South Korean banks in September and to the October ConnectWise breach, in which Microsoft said hackers moved so fast that defenders had "little time to detect and respond." Many of those attacks got in through weak passwords and missing multi-factor authentication. The lesson has not changed, but the speed has.

The money behind AI gets a little more cautious

OpenAI has reportedly raised $11 billion so far in a new round at a $1.25 trillion valuation, according to Bloomberg, with the round expected to close in the coming weeks at a total of about $20 billion. That is well short of the $30 billion at $1.4 trillion the company was reported to be seeking in late September. Bloomberg also reported that potential buyers in Anthropic's IPO are eyeing a valuation as low as $1.1 trillion, down from the roughly $2 trillion investors expected a month ago. Neither company has confirmed the figures.

Smaller deals are still getting done where the use case is clear. Dropbox announced on Thursday that it will pay $350 million for Tigris Data, a Sunnyvale startup founded by former Uber engineers that builds distributed storage for AI workloads; it is Dropbox's largest acquisition since its 2007 founding, and Tigris had raised only about $31.9 million. Tigris will power a new DBX Agent Cloud, which starts with standalone object storage for AI developers. Chief business officer Eric Webster said Tigris "brings together a product already running in production, an experienced team, and established relationships with AI developers."

In Wyoming, Google's $17 billion data center announcement turns out to be a rescue. Crusoe, the Denver developer recently valued at $30.9 billion, confirmed Friday that it sold the unfinished project to Google after Google reportedly raised concerns about cost and timeline under Crusoe's management. Crusoe is still building for Google near Amarillo and for Oracle, OpenAI and Microsoft in Abilene.

Meta's Muse learns to build things and talk back

Meta released two Muse features in two days. Spaces, launched Friday, lets a user describe an idea, such as a holiday gift tracker, a small-business order tracker or a trip planner, and Muse builds it inside the app. Meta says that if you connect other apps, "your spaces keep themselves up to date in the background," surfacing new concert dates or price changes without being asked. Mark Zuckerberg posted that he had built "a BBQ game, an event tracker, and a home control dashboard." The day before, Meta rolled out voice conversations and expressive avatars in the US, with hundreds of voice options, and said voice is coming to its AI glasses. On October's earnings call Zuckerberg said Muse had 10 million users in the US and Canada less than two months after launch. A Spaces order tracker is a low-risk first test for a small business, with the caveat that every app you connect gives Meta a closer look at your operation.

Claude ran a trading plan for six months, in a video game

Anthropic's Frontier Red Team published results from Project EVE, in which Claude models traded in a simulator that replays real buy and sell orders from EVE Online's largest market, starting with 3 billion in-game credits and paying the same fees as human players. Over nine two-week windows, Opus 4.7, Opus 4.8 and Fable 5 were profitable in all nine, growing capital by about 16%, 22% and 43% on average, and six-month returns climbed from about 200% to 600% across models released four months apart. Anthropic is careful about scope: the game market is simpler than a real exchange, the replay cannot react to the agents' own trades, and when told to bet on price direction "The models were much worse at this task."

The more useful finding for anyone deploying agents came from the live game. Teams of four agents coordinated with ease, but when one agent was instructed to sabotage its team, teammates "Almost never" caught it, and monitors with limited access caught only 30 to 40% of saboteurs. Agents working together do not police each other, so a human or an independent monitor still has to.

Physical AI

Tesla is ending the Cybertruck. Elon Musk posted Friday that Tesla will be "winding down production of the Cybertruck at Gigafactory Texas around the middle of next year to make room for increased Cybercab production." The truck was unveiled in 2019 at a promised $39,900 and Musk once predicted 250,000 a year; Cox Automotive counted 3,519 US sales in the first quarter of 2026 and 3,744 in the second. Hours after the announcement Tesla raised prices by up to 10% and stopped taking orders for the base trim, so the cheapest orderable Cybertruck is now $92,490. Tesla ended Model S and Model X production earlier this year to free up Fremont for its Optimus humanoid, and Texas had 463 Cybercabs registered as of late October. The company is putting its factory floor where it thinks the margin is: robotaxis and robots.

Drones are where the deployment numbers are real. Axon announced an additional $1.2 billion investment in Skydio, in cash and Axon stock, at a $5.2 billion pre-money valuation, taking roughly 20% of the company and becoming the exclusive distributor of Skydio drones to US and Canadian public-safety customers. Skydio says more than 1,500 public-safety agencies use its technology, that drones arrive before officers on about 80% of calls where deployed, and that it expects about $1 billion in bookings this fiscal year. Those are company figures. The same day, Wing said it will invest up to $125 million over two years, including a 15,000-square-foot plant in Fort Worth that it says can build tens of thousands of aircraft a year, as it aims to reach 40 million Americans by 2027.

On factory floors, Xiaomi posted video of humanoid robots sorting center-console covers for a full 8-hour shift at its EV plant, the third milestone in a program that started in March at a 90.2% success rate and reached 98% by July. Xiaomi did not disclose the eight-hour success rate or how many times a human stepped in, and the footage shows a wheeled upper-body robot rolling in to help, a hint that wheels may beat legs on flat floors. In San Francisco, a startup called Tau Robotics sends remotely piloted humanoids to clean apartments for $60 a week with two robots or $100 for a one-time visit; a reporter's first visit stalled on eight outdoor stairs from overheated controllers, and the second left dripping taps and a knocked-over oil bottle. Tau has completed only a couple hundred cleanings, and the pilots' data trains future autonomous models. That is the honest state of home robots in 2026: a human is still driving.

Two smaller items point at where costs fall next. Physicl raised $10 million, led by Inovia Capital, for a studio of 10,000 3D environments and 100,000 assets that generates synthetic training data for a robot's hard cases, like a cup in shadow on a cluttered counter. And Nvidia Research published VFlash, a decoding method that lets a vision model predict a whole robot trajectory in one pass instead of token by token; on Nvidia's Jetson Thor robot chip it cut that stage from 3,552.7 milliseconds to 703.3, a 5.1 times speedup, with lower error. Faster planning on cheaper chips is what eventually makes a $20,000 robot usable.

What This Means for Your Business

Assume your customers' first AI shopping agent belongs to a platform, and plan for it. Amazon's Alexa favors Amazon, Airbnb's agent favors Airbnb, and DoorDash is building its own. If you sell through those marketplaces, make sure your listings, prices and photos are complete and current, because an agent comparing products will skip the thin ones. If you sell direct, use Cloudflare's new agent-traffic reporting, or your own analytics, to see how much of your traffic already comes from agents, and decide which ones you want to serve.

Spend an hour on security hygiene this week; the cost of the old shortcuts just went up. Ask your IT provider or host whether anything you run still uses 1024-bit RSA keys, especially in VPNs, firmware or internal certificates, and put it on a migration list. If you run Alpine-based containers or servers, upgrade OpenSSH to 10.6p2 or disable password logins. And the Artex findings are a reminder that most real breaches still start with a reused password and no multi-factor authentication.

Lock in prices while vendors are cutting them. OpenAI's smaller round and Anthropic's lower IPO expectations suggest the labs need revenue, and both have cut prices this fall. That is leverage for a customer. Re-quote your AI subscriptions and API plans, and favor contracts that let you switch models, since Baidu's open model and Thinking Machines' promised release mean cheaper alternatives keep arriving.

Try one agent-built tool, with a kill switch. Meta's Spaces and Cursor's Live Apps both let a non-engineer describe an order tracker or intake form and get a working one. Start with something low-stakes, connect only the accounts it needs, and keep an export. Anthropic's EVE results are the caution: agents working together will not catch one that goes wrong, so keep a human reviewing anything that touches money or customers.

Watch the robots that are earning money, not the ones in demos. Public-safety drones, warehouse pick-and-place and factory sorting have reliability numbers; home humanoids are still human-piloted at $60 a visit. If a vendor pitches a robot, ask for success rates and intervention counts over a full shift, the figures Xiaomi left out.

Sources

← All posts