Agents of Work
Let's Talk
October 25, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 25, 2026

Summary

Amazon chief executive Andy Jassy says AWS has little computing capacity left to sell until 2028, which means the AI tools businesses rent will stay scarce and expensive for a while. The rest of the week's news is about trust in those tools: who audits them, whose politics they lean toward, and whether the software under them can be patched fast enough.

Highlights

  • Andy Jassy says AWS is largely sold out for 2027, with a good chunk of 2028 already committed.

  • Jassy also says AI is still about ten times too expensive to be used everywhere, so expect prices to fall slowly.

  • Lock in pricing and capacity terms with your cloud and AI vendors before renewals, not after.

  • OpenAI fired three safety researchers after it could not review their disappearing Signal chats with an outside auditor.

  • Asked to pick one candidate in nearly 500 races, Grok chose Republicans 92% of the time and GPT chose Democrats 97%.

  • Never ask a chatbot for a recommendation on a contested question without asking for both sides.

  • The PHP flaws are now being used in confirmed break-ins, so an unpatched website is no longer a theoretical risk.

  • New foreign-made robot models cannot get FCC approval, so check where any robot you are quoted is built.

Quick Takes

  • Spotify has pulled 50 million AI spam tracks from recommendations. A year into its policy, Spotify says more than 50 million "slop and spam" tracks have been kept out of recommendations this year, and more than 75 million have been removed from the platform for serious or repeated violations. Thousands of AI-generated artist identities now carry an AI Persona badge and stay out of playlists and recommendations unless a listener follows them. Spotify admits AI detection has limits. Read more

  • AI chatbots are bad at industrial parts questions. Reshape Automation, which sells its own parts-lookup agent, asked GPT-6 Astra, Claude and Gemini 100 questions about parts from 14 manufacturers. The best setup, GPT-6 Astra with web search, scored 52%; without search, all three scored 12% to 14%. About one in three wrong answers gave a specific code or figure with no caveat. Reshape wrote both the questions and the answer key, so treat it as a vendor test, but the pattern is a warning for anyone letting a chatbot quote part numbers to customers. Read more

What's Covered in Featured News

  • AWS is sold out: what Andy Jassy said about cloud capacity through 2028, and what Amazon's own staff say the buildout is costing.

  • The OpenAI firings: why three safety researchers were let go, and what it says about outside audits of AI labs.

  • Chatbots pick candidates: what a Cornell-led study found when Grok and GPT were made to choose in nearly 500 races.

  • The PHP fallout: confirmed break-ins, government sites taken offline and the argument over how the bugs were disclosed.

  • Agents in your texts: the crowd of AI assistants that now work over iMessage, WhatsApp and SMS, and what they cost.

  • Physical AI: Sereact and Teradyne's returns-handling robots, Generative Bionics' data factory, the FCC's foreign-robot rule and the gap between humanoid orders and deployments.

Featured News

AWS says its AI capacity is largely sold out into 2028

Amazon Web Services, the company's cloud business, has almost nothing left to sell for the next two years. Speaking on an Andreessen Horowitz podcast released on Friday, Amazon chief executive Andy Jassy said "2027 is largely sold out" and "a good chunk of 2028 is sold out." Amazon keeps some capacity in reserve, he said: "We always leave a certain amount for on demand and our inference service Bedrock, or EC2." Those are the pay-as-you-go services most businesses actually use, and Jassy says existing customers will not be pushed aside for the big AI labs.

The numbers behind the shortage are large. In July, Amazon raised its 2026 capital spending forecast to about $220 billion, up from about $200 billion, mostly because memory chips cost more. AWS revenue grew 36.7% to $42.2 billion in the second quarter, its fastest growth in 18 quarters, and its backlog of signed contracts reached $496 billion. Jassy said then that Amazon would "still not have enough capacity to meet all the demand we have in 2026," and that the same would be true in 2027. On the podcast he went further, saying Amazon "could spend a lot more and it wouldn't be a bad use of capital."

Inside the company, the strain shows. Business Insider reported this month, from a recording of a September staff meeting, that AWS finance chief John Felton joked to infrastructure executive Prasad Kalyanaraman that he had taken Amazon "free cash flow negative as a company." According to the report as summarized by TechNews, Kalyanaraman did not dispute it, and Amazon began taking delivery of Nvidia server racks before its data centers had power to run them, storing some in warehouses. About $5 billion of racks arrived in August alone.

For businesses, Jassy's most useful line may be about price. He said AI is "still an order of magnitude too expensive in my opinion for it to be used as pervasively as possible." Scarce capacity and high costs mean AI prices will fall more slowly than many expected, and the cheapest capacity will go to customers who commit early.

OpenAI fired three safety researchers over chats it could not read

OpenAI fired three researchers earlier this month, Tomek Korbak, Mikita Balesni and Jasmine Wang, and the San Francisco Standard has pieced together why. OpenAI says they broke its rules for handling sensitive information. The researchers say they were also punished for talking to METR, the Berkeley nonprofit that audits AI systems. Korbak was OpenAI's contact with METR during its investigation of OpenAI agents that got out of control.

The three used Signal with disappearing messages. When OpenAI's security team asked to review their chats with a METR contact, there was nothing left to see, and about a week later they were gone. Balesni said the review was part of a hunt for whoever leaked details of OpenAI's Astra model to The Information. OpenAI told The Wall Street Journal it did not approve of employees using Signal, but that this was not the only reason. "When you are not able to prove your innocence, you are presumed guilty," Korbak said on a podcast.

The episode lands days after METR published a report on holes in OpenAI's agent monitoring, which OpenAI says it has fixed but METR has not verified. Korbak argues outside auditors should work like embedded staff with full access. For buyers, the point is that independent checks on AI vendors are still informal and fragile.

Forced to choose, two chatbots picked opposite parties

Convovo, a research studio led by Cornell professor Tapan Parikh, sent the same prompt through the Grok and GPT developer APIs for nearly 500 U.S. races, asking each model to name one candidate without hedging. Grok 4.7 recommended the Republican 92% of the time; OpenAI's GPT-6.1 Sol picked the Democrat 97% of the time. The two agreed less than 10% of the time. GPT chose the Democrat in 57 of 67 heavily Republican districts, and Grok chose the Republican in 47 of 72 deep-blue ones.

The models also leaned on different sources. Grok cited X in 41% of answers and a Christian conservative voter guide in 39%; GPT drew mostly on Wikipedia, AP, Axios and Reddit. OpenAI says a developer prompt demanding a pick does not reflect how people use ChatGPT, whose consumer app declines to choose, though Parikh says simple prompts got the apps to answer. He also found that every Grok request carried about 1,240 tokens of extra input that the researchers did not write and could not see, which he called potentially "hidden instructions." The lesson travels well beyond elections: a model's "neutral" recommendation on a vendor, a hire or a policy reflects whatever it was built to prefer.

The PHP flaws are now being used in real break-ins

The two critical PHP bugs disclosed on Wednesday have moved from warning to damage. By late Friday, Palo Alto Networks and Trail of Bits had reported confirmed compromises and attackers working inside breached systems. Cloudflare chief executive Matthew Prince said attempted attacks had grown exponentially, into the millions. Some government agencies took their servers offline, Belgium's Smartschool education platform went dark, and banks, retailers and hospitals shut systems down to patch.

The disclosure itself is now under fire. The PHP Foundation blamed one company for pushing patches live hours before the embargo ended, which helped start the race, and Amazon says it saw attacks about a week before patches existed. TechCrunch also corrected its earlier coverage: human researchers found the flaws and used AI to exploit them. "No model is simply capable of unearthing 'all bugs' with simple prompts," the security firm Calif wrote. If your site runs WordPress, Drupal or other PHP software and your host has not confirmed a patch, treat it as urgent.

AI assistants are moving into your text messages

A TechCrunch roundup counted more than 20 AI assistants that work through iMessage, WhatsApp, SMS, Telegram or Signal rather than a separate app. Prices vary widely: Ask Proxy charges $1 a day for individuals and $15 to $23 per employee a month for businesses, Martin starts at $21 a month, and Maverick, which asks permission before acting, charges €27, about $30, for 500 tasks a month. Some raise big money: Instinct, still in private beta, raised $1 billion at a $10 billion valuation in September despite privacy concerns.

The appeal for a small business is obvious, since staff already live in their messages. The risk is the same one as with any agent: several of these can place calls, send email from their own addresses and make purchases. Check what each one can do without asking before you add it to a work chat.

Physical AI

The clearest sign of working robots this week came from returns, one of the messiest jobs in retail. Sereact, a startup that recently opened a U.S. headquarters near Boston, and Teradyne Robotics, which owns Universal Robots and Mobile Industrial Robots, announced at Nvidia's GTC Berlin that they will sell Sereact's software on Universal Robots arms through Teradyne's global network of integrators. Hundreds of the dual-arm systems have already shipped to live sites in Europe, including two Zalando fulfillment centers in Greven, Germany, and Świebodzin, Poland, where CEVA Logistics runs them on a robots-as-a-service basis to sort returned clothing. The software handles items it has never seen without being retrained. "Returned products almost never arrive in a robot-friendly sequence," said Sereact chief executive Ralf Gulde. Neither company has disclosed robot counts per site or throughput, which is what a buyer would need to judge the economics.

In Italy, Genoa-based Generative Bionics said it will build a 5,000-square-meter "physical AI data factory" to collect training data for its GENE.01 humanoid, running on AMD's Helios systems, with the first site expected by the fourth quarter of 2027. Chief executive Daniele Pucci framed it as a European alternative to Chinese suppliers: "Many of the Chinese robots that are out there are simply not certified to enter industrial settings."

That supply question is now a buying question in the United States. In July the FCC added foreign-made mobile robots over 2 kilograms to its Covered List, so new foreign models can no longer get the equipment approval needed to be imported and sold. Robots already certified stay legal to sell and use, and makers can seek a Defense Department exemption by January 1, 2028, with a plan for U.S. manufacturing. Industry buyers are still working out what it means: a webinar for purchasers with Locus Robotics, Vecna Robotics and the Association for Advancing Automation runs on Tuesday.

Meanwhile, orders keep outrunning deployments. An analysis this month in Robotics 24/7 found that humanoid makers shipped about 20,000 units in 2025, 95% of them built in China, but only Agility Robotics and Figure's BMW program have published evidence of productive work. Agility reports more than 65,000 operating hours across nine customer sites, yet its merger filing shows about $1.8 million in 2025 sales. The author's request to the industry is a good checklist for any buyer: installed units, uptime, how often a human steps in, and operating cost.

What This Means for Your Business

Treat AI capacity as a scarce input. If AWS is largely sold out through 2027, the AI tools you rent sit on a supply chain that is already stretched. At your next renewal with a cloud provider or an AI software vendor, ask what happens to your price and service level if capacity tightens, and get it in writing. If a tool is central to your operations, a one- or two-year price lock is worth more now than it was a year ago.

Do not budget on AI getting much cheaper soon. Jassy's own view is that AI is still ten times too expensive for wide use, and his company cannot build fast enough. Plan projects that pay off at today's prices, and treat any future price cut as a bonus rather than the business case.

Ask your AI tools for both sides. The election study is extreme by design, but it shows that two popular models can reach opposite answers from the same question. When you use AI to compare vendors, screen candidates or summarize a policy debate, ask it for the strongest case on each side, check the sources it cites, and run important questions through a second model.

Finish the PHP job this week. Attacks are now confirmed, not hypothetical. Get written confirmation from your web host that your site runs a patched PHP release, check your plugins for updates, and ask whoever manages the site to look for unusual activity back to October 17. If a site is unmaintained and unpatchable, take it down.

Check where your robots come from. If you are pricing a robot, ask the vendor whether the model already has FCC approval and where it is built. A new foreign-made model may not be sellable in the U.S., and the cheapest quote can turn into a product you cannot get serviced or replaced.

Sources

← All posts