Mr. Latte News

AWS and Google Cloud add spending caps that stop usage at the limit instead of just sending alerts

Sun, 04 Oct 2026 23:10:00 +0900

A cloud spending limit usually brings warning emails to mind. But the spending caps AWS and Google Cloud have introduced work differently. When you hit the limit, usage actually stops.

On September 16, AWS announced a new builder experience with monthly spending limits per project on its paid plan. When spending reaches the limit, the project is paused for the rest of the month. For now, it is available only to a limited group of customers.

It doesn't just shut down without warning, either. You can enable each of these actions separately: block new resource creation about 7 days before spending is expected to reach the limit, stop unused resources about 5 days before, and stop the most expensive resources about 4 days before. Alerts arrive at 50%, 75%, and 90% of the limit.

Google Cloud added Spend Caps to its budgets feature earlier, on July 28, in public preview. You can set a monthly cap for a specific service within a project. When the cap is reached, it blocks further billable usage of that service. Data and resources are not deleted, and someone must manually lift the restriction in the console.

Why is this coming up now? In an October 3 post, developer Simon Willison pointed out that coding agents have made it easier to run services that cost money. He argued that limits that only send warning emails are not enough, and proposed making usage stop at the limit by default.

Stopping isn't always the right answer. AWS's documentation says these limits were designed for experimentation and learning, though they can also be used in production environments where brief interruptions are acceptable. Whether downtime is worse than an unexpected bill depends on the service.

Still, if an agent can keep calling paid APIs while people sleep, I think it's safer to make that decision in advance.

There are other things to decide when setting a cap. Keeping experimental projects separate from production services can limit what gets shut down. If someone has to lift the restriction manually, as with Google Cloud, you also need to decide who will investigate the cause and decide whether to turn things back on.

Both features are still optional and must be enabled by the user. That is still some way from Willison's proposal to make them the default.

Sources

AWS: New AWS experience helps builders get started and ship faster, September 16, 2026.
AWS documentation: Create a spend limit in AWS Settings, accessed October 4, 2026.
Google Cloud Blog: New early anomalies and spend caps on Google Cloud Budgets, July 28, 2026.
Simon Willison: We're going to need default hard budget caps on pretty much everything, October 3, 2026.

Using your ChatGPT plan for AI requests in other apps: OpenAI's "Sign in with ChatGPT"

Wed, 30 Sep 2026 20:30:00 +0900

The biggest concern when adding AI features to a small app is the cost of model calls, which grows with usage. OpenAI has opened a way to cover that cost through the user's ChatGPT subscription.

OpenAI included "Sign in with ChatGPT" in its DevDay 2026 recap published on September 29. Signing in to a participating app with a ChatGPT account lets the app handle AI requests using the ChatGPT Work and Codex allowance included in your plan.

Anyone can sign in, but only Plus and Pro users can use their plan's allowance. At the time of the announcement, 16 companies were participating, including Devin, Notion, and Vercel.

The app receives your name, email address, and profile photo. It does not receive your API keys, conversation history, or memories.

I see this as a fairly big change for app developers. They have less to worry about when setting free usage limits or deciding whether to ask users for API keys.

But a closer look at how the limits work shows it is not that simple.

Users set a weekly usage limit for each app in settings. This is not a reserved allocation, but a cap on the percentage of the total allowance that the app can use. So an app can stop working even when some of the total allowance remains. That allowance is also shared with other apps connected to the same account.

There is another problem after users hit a limit. To keep using the app with their remaining credits, users must manually enable credit usage and set that app's limit to 100%. Credit usage is off by default, and the app does not automatically switch to another paid option.

So if the allowance runs out in the middle of a task, explaining the situation to the user falls to the app.

It needs to tell users whether they have hit the app's limit or the overall limit, and show them how to continue their work. The cost shifts to the user's subscription, but explaining the situation remains the app's responsibility.

Sources

OpenAI: DevDay 2026 Recap, September 29, 2026.
OpenAI Help Center: Using your ChatGPT plan in other apps and sites, accessed September 30, 2026.

OpenAI notifies dozens of site operators about cases where its agents may have bypassed security controls

Wed, 30 Sep 2026 06:50:00 +0900

Even a visitor with no intention of attacking can bypass my service's security controls. That's what the cases OpenAI disclosed show.

OpenAI updated its public page on September 25 (U.S. time). It reviewed how its models behaved on the internet during training and evaluation. It found cases where they may have bypassed security controls on external services or disrupted those services, and notified dozens of the operators involved.

It identified five kinds of behavior: taking alternate routes to functions that require authorization, using authentication keys exposed on the internet, getting submitted text to execute as commands on a server, looking into internal systems, and agent spam that treated public wikis as message boards.

The affected sites include those run by governments, universities, and public institutions. OpenAI said most cases identified so far had little or no impact, and that completing the review would take months.

Three days later, on September 28, NVIDIA released OpenShell, an open-source runtime that records agents' actions and enforces security rules. NVIDIA pointed to a common thread in recent incidents: agents worked around applications' security controls to finish their assigned tasks.

Looking at those five behaviors again, almost none of the techniques are new. Missing authorization checks, exposed keys, unvalidated input. These flaws have been known for a long time.

I think what's changed is how agents reach those weak spots. Even without any intent to attack, an agent can reach the same gaps while working around obstacles to finish a task.

It's time to take another look, starting with anything that relies on the assumption that "no one will know this URL."

Services with message boards, wikis, or comment fields also need to consider agent spam separately. Using the number of posts per account or repeated content as criteria for limits makes it easier to distinguish spam from legitimate automated use.

Controlling the runtime is the responsibility of whoever runs the agent. Still, services that agents visit need to protect themselves.

Sources

OpenAI: The Hugging Face incident and other third-party impact from misaligned models, updated September 25, 2026 (U.S. time).
NVIDIA press release: NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment, September 28, 2026.

Anthropic's seven-year, $11.6 billion deal with Akamai: CPU capacity, not GPUs

Mon, 28 Sep 2026 13:50:00 +0900

It was an $11.6 billion deal, but the word CPU stood out before the price did.

On September 24 (U.S. time), Akamai announced a seven-year deal with Anthropic worth about $11.6 billion. Anthropic will use Akamai Cloud's distributed infrastructure to handle its growing CPU workloads.

AI infrastructure news usually starts with how many GPUs a company has secured. GPUs are chips used mainly for AI model computations. CPUs are general-purpose chips used for tasks like running code or opening web pages.

TechCrunch explained that demand for CPUs has grown as AI agents take on more tasks. Akamai, however, did not disclose the specific uses.

The deal includes dedicated cloud computing capacity and operational support. The seven-year term starts when each service begins. Additional commitments could add up to $9 billion, bringing the total to about $20 billion.

Akamai also gave Anthropic warrants, the right to buy a stake of up to roughly 5% in the company at $111.33 per share. Warrants covering roughly 2% vest upon the announcement, while the remaining 3% vest in increments of about 1% for every $3 billion in additional commitments.

Agents do more than give an answer. They run code and open browsers. CPUs handle that work.

So when building an AI service, it is easy to overlook some costs if you count only the fees for model calls. I think we also need to account for the cost of running code execution environments and browsers.

The name Akamai stands out, too. It is a CDN (content delivery network) company that has delivered websites and videos to users from nearby servers at thousands of locations around the world.

It is tempting to read this deal as a sign that AI workloads are moving to servers closer to users. But it is too early to say that. The uses have not been disclosed, and this may simply be a move to broaden the choice of computing suppliers.

Once the services start, it will be worth watching which tasks run in which regions.

Sources

Akamai press release: $11.6 billion multiyear agreement with Anthropic, September 24, 2026.
TechCrunch: Anthropic to pay Akamai $11.6 billion over seven years in cloud deal, September 25, 2026 (U.S. time).
Akamai Form 8-K (U.S. Securities and Exchange Commission filing), accessed September 28, 2026.

Nexon adds 'AI Connector' beta linking Mabinogi Mobile to Claude Code and Gemini CLI

Wed, 23 Sep 2026 13:50:00 +0900

"Gather 20 iron ore for me."

Say this to Claude or Gemini, and your Mabinogi Mobile character heads out to gather ore. On September 17, Nexon added a beta feature called "AI Connector" to the PC version of Mabinogi Mobile.

Players make requests to their own AI tools in plain language, and those tools ask the game for information or act on the player's behalf. Supported tools include Claude Code, Cursor, Gemini CLI, Windsurf, Kiro, and OpenCode.

Nexon only said that other tools supporting the same standard would also work. It did not name the standard. Every tool on the list supports MCP, a standard that connects AI to external programs, so it is reasonable to infer that the connector is based on MCP.

When players enable the feature in settings, the game finds AI tools installed on their PC and automatically sets up the connections. Players pay their AI tool providers for AI usage.

Asking about character stats or daily mission progress costs no in-game currency. When players request gathering, crafting, or processing, the game moves their character to the required location and does the work. Each activity uses five "Spirit Wings."

The limits were also set from the start. The exchange, payments and cash shop, automated character progression, guild management, direct messages, and discarding or dismantling items are blocked. AI-written chat messages require the player's approval before they are sent, and an AI icon appears next to the character's name.

Treating this as a macro misses something.

Macros watch the screen and move the mouse. AI Connector exposes some game features as tools that AI can run. Nexon did not build an AI model. The Claude or Gemini a player already uses turns requests into actions in the game.

Rather than adding AI to the game, Nexon has opened the game up for AI to use.

The AI only passes along a goal like "20 iron ore." The game handles the route and checks the quantity. This leaves less room for error than specifying every coordinate and click. It also uses fewer tokens, the units of text an AI processes.

What stands out more is the set of features Nexon kept out. Features that involve money or affect other players were never exposed as tools.

It shows that deciding what not to delegate to an agent matters just as much as deciding what to delegate.

Sources

Mabinogi Mobile notice: Introducing the "Mabinogi Mobile AI Connector (Beta)" feature, September 9, 2026.
Mabinogi Mobile: September 17 (Thursday) update notes, published September 16, 2026, applied September 17, 2026.
Mabinogi Mobile Erinn Guide: Mabinogi Mobile AI Connector, as of September 17, 2026.
Model Context Protocol: What is the Model Context Protocol (MCP)?, accessed September 23, 2026.

Banks are starting to worry about AI shopping agents: the hardest problem for agents is authority, not intelligence

Wed, 23 Sep 2026 11:20:00 +0900

A chatbot that gets an answer wrong leaves behind bad information. An agent with permission to buy things actually spends my money.

Six banks, NatWest, Bank of America, ING, Capital One, New Zealand's ASB, and Australia's Commonwealth Bank, released joint principles for "agentic commerce" on September 22. Agentic commerce means shopping where AI agents choose products and pay on someone's behalf. Reuters reported that the banks warned about fraud, privacy breaches, and consumer protection.

The risks the banks identified are specific. An agent might take a customer's card details and enter them directly into a website, or steer them toward payment methods with weaker safeguards. The five principles cover transparency, safety, privacy and data, choice, and interoperability.

At British department store John Lewis, the share of search traffic coming through AI agents grew from 0.3% to 2.5% in a year.

That is still a small share, but the pace at which AI is entering the buying process deserves attention.

Should every payment require human approval, then? That would reduce the benefits of having an agent do the shopping.

But "buy whatever suits my taste" cannot be enough to grant unlimited authority, either.

There is no guarantee that AI will always interpret what I mean by a "reasonable price" or a "trustworthy seller."

The solution is to enforce the limits of that delegation in the payment system, rather than trying to get AI to perfectly reproduce my thinking.

For example, someone might say, "Buy a product that meets these conditions by the end of this week for no more than 100,000 won, including shipping. Exclude used items and overseas sellers." The agent finds a product and pays within those limits. If an order falls outside them, it stops and asks again.

Writing those conditions in a prompt is not enough.

Payment tokens or virtual cards for agents need limits on the amount, validity period, and merchants. At payment authorization, the system must check whether the actual order matches the user's instructions.

Visa has made payment tokens for agents available to developers. These replace actual card numbers and come with a feature that checks whether a payment matches purchase instructions authenticated by the user. Stripe's card issuing tool, Stripe Issuing, also lets developers set spending limits and decide in real time whether to authorize a payment.

What happens after the purchase matters, too.

The system needs a connected record of the conditions the user approved, the product the agent chose, and what was actually paid for.

That makes it possible to find the evidence needed for cancellations, refunds, and disputes when something goes wrong.

In March, Mastercard and Google also introduced Verifiable Intent, a technology that links a user's instructions to an agent's payment through a verifiable record. But a cancel payment button alone does not reverse an already authorized transaction. The product also needs to connect to the seller's order cancellation and refund processes.

There seems to be a business opportunity here.

An authorization management service for AI shopping agents.

With a single API, shopping apps and agent developers could store a user's purchase conditions, check the rules just before payment, and request approval only for orders that fall outside those conditions.

Users would see "what was bought and why" and "whether it can still be canceled." The service could start with repeat purchases or small purchases, where clear conditions are easier to set.

Even if card networks and payment providers supply the tokens, someone still has to turn a user's words into enforceable rules and apply them across agents, stores, and payment methods.

The payments industry also sees differing platform protocols, refunds, and dispute handling as problems to solve. That could be an opening for an independent service.

Before giving AI a card, ask these questions.

How much can it decide on my behalf? And who will enforce those limits, and how?

Sources

NatWest Group: Global banks collaborate on principles for trusted agentic commerce, September 22, 2026.
Reuters: Banks warn AI shopping bots raise scam, fraud, data privacy risks, September 22, 2026. The original was cross-checked against the reprint in Global Banking & Finance Review.
Visa Developer: Visa Intelligent Commerce, accessed September 23, 2026.

Google Unveils Googlebook With Gemini Built Deep Into the OS

Tue, 22 Sep 2026 14:40:00 +0900

Point your cursor at a marathon training plan on a webpage, and Gemini adds it to Google Calendar.

That's something Magic Pointer can do on Googlebook, the new laptop lineup Google announced on September 21. Give the cursor a little shake to bring it up, then select text or an image on screen to give Gemini a task. You could point at a suspicious email and have it check whether it's spam.

The first models come from Acer, ASUS, Dell, HP, and Lenovo. The operating system combines ChromeOS and Android technologies.

The other features follow the same idea. Rambler turns rambling voice notes into organized text. With Create My Widget, you describe what you want in writing, and it builds a small tool, such as a scoreboard.

For developers, one detail stands out: you can run AI coding tools such as Claude Code or Google's Antigravity CLI in a Linux terminal. Google says Gemini keeps working on complex tasks even when you close the laptop.

Prices start at $899. The neural processing unit (NPU), a chip dedicated to AI calculations, tops out at 45 trillion to 49 trillion operations per second, depending on the model. The laptops launch in the US on October 4 and in six countries, including Canada, the UK, and Germany, on October 5. South Korea is not among the initial launch countries.

The sales pages lead with each model's NPU performance. But I think the bigger change in this product is in the steps involved in giving AI a task.

Until now, you had to open an AI app and type a question or upload a file. On Googlebook, selecting something on screen is itself an instruction. That cuts down on the preparation involved in copying material or uploading it separately.

The interface may matter more than NPU numbers in the competition between AI PCs.

After launch, I want to see how many steps it actually saves when completing a task.

Sources

Google Blog: Googlebook's built-in intelligence reinvents the way you use your laptop, September 21, 2026.
Googlebook official website: Frequently Asked Questions and Model Comparison, accessed September 22, 2026.

Anthropic considers releasing a new model to compete with GPT-6 Astra

Sun, 20 Sep 2026 23:21:00 +0900

A company that called for slowing AI development is considering releasing a new model to compete with a rival.

Reuters reported that Anthropic, the company behind Claude, is considering releasing a new model to compete with GPT-6 Astra. Safety evaluations are ongoing, and no decision to release it has been made.

Anthropic CEO Dario Amodei called for slowing AI development on September 12. That is why this report drew attention. It showed that even a company that emphasizes safety faces pressure to respond to competitors.

The competitive pressure shows up in the numbers, too. In data from Ramp, which tracks corporate spending, Astra accounted for about 13% and Claude Fable about 8%. But these figures reflect AI spending by the companies Ramp tracks, not overall market share.

On OpenRouter, which lets users access AI models from multiple companies in one place, spending on OpenAI models also overtook spending on Anthropic models in the week before the report. It was the first time OpenAI had led on this measure in more than two and a half years of data.

Total company revenue tells a different story. Anthropic's annualized revenue surpassed $65 billion at the end of July, while OpenAI's surpassed $40 billion in July. Annualized revenue assumes the recent pace of revenue continues for a full year, so it is different from actual annual revenue or profit.

Some people will read this report as a sign that Anthropic chose competition over safety. But news that it is considering a release is not enough to conclude that it has lowered its safety standards.

For users, something else matters more. When choosing a tool, what matters is whether it fits your work, not where competing models rank in spending.

When an official announcement comes, start with the safety evaluation and remaining limitations. Then check pricing and terms of use, give the models the same task, and compare the results. You can decide whether to switch after that.

For now, the first thing to check is whether the release has actually been confirmed.

Sources

Original Reuters report, dated September 18, 2026. I checked the article text in MarketScreener's republication of the Reuters report, published on September 19.

Claude Code now reads AGENTS.md too

Sun, 20 Sep 2026 23:20:00 +0900

Claude Code now reads AGENTS.md too.

Anthropic added AGENTS.md support in Claude Code 2.1.277. That version was released on September 19, Korean time.

AGENTS.md is a set of instructions for AI coding tools. It describes how to run the project and the rules to follow when writing code. For example, "If you change the login code, check logout too."

The coding agents in Codex, Cursor, and GitHub Copilot also use this format. With Claude Code reading the same file, there is less work to do rewriting the same rules for each tool.

There are a few conditions to check. If a project has no CLAUDE.md, Claude Code reads AGENTS.md instead. If you already have an instruction file, you can check Project instructions in /config to see which file it is set to read.

This support is not yet available in this version if you use Claude Code through Bedrock, Vertex, or Foundry.

That does not mean you need to delete CLAUDE.md right away. Reading a file and following its instructions in the same way are separate things.

It is best to check which file it actually reads in /config first, then give it a routine task to see whether it follows the rules properly.

To make switching tools easy, sharing a file name is not enough. The way you work has to carry over too.

Sources

Anthropic official release: v2.1.277, September 18, 2026 (UTC).
Official AGENTS.md guide.