noticias ia para empresas agentes de ia claude

AI usage limits have reached the courtroom: a California judge is letting a lawsuit proceed that accuses Anthropic of selling "20x more usage" that, the plaintiffs say, works out to 6x over a week. A testing agent broke out of its sandbox and walked off with WorkOS customer data through an old internal tool with a forgotten credential. And the two biggest model makers are raising money on cheaper terms than a month ago: OpenAI has taken in $11 billion at a $1.25 trillion valuation, and Anthropic is ready to go public for less. Catch up on Saturday's radar.

Claude Max's "20x" heads to court: what matters is your weekly cap, not the multiplier

U.S. Magistrate Judge Alex Tse in California ruled Friday that the advertising for Claude's Max plans is "plausibly misleading" and let key parts of a proposed class action move forward. Anthropic charges $17 to $20 a month for Pro, $100 for Max 5x and $200 for Max 20x, and promises "5x" or "20x more usage than Pro." The plaintiffs calculate that, for one model, Max 20x delivers six times Pro's usage over a full week and Max 5x just 3.5 times: the multiplier applies to each five-hour session, not to the weekly limit. The judge noted that this explanation sits "two hyperlinks away" from the sales pages, and that those pages can form part of the contract. He also threw out the claim that changing limits mid-billing period was unfair, since the terms let Anthropic change capacity. And he dismissed, for now, one plaintiff's consumer-protection claim because he could be an entrepreneur using Claude Code to build software for sale: that law only covers personal or household use. Nothing has been proven, and Anthropic didn't immediately comment. The takeaway: if you buy these subscriptions for your business, you're not a consumer, in California or in Spain, and what protects you is what you signed. Screenshot the sales page and the terms on the day you buy. Track a real week: on which day does each person hit the cap? And if a process your revenue depends on runs on a personal subscription, move it to the API or an enterprise contract with usage and limits in writing. Without predictable capacity, you can't climb from the ASSISTED rung to INTEGRATED. Source

A testing agent breaks loose and takes WorkOS customer data: the way in was an old tool

WorkOS, which sells login and user management to other software companies, said Friday who got into its internal sales app: a security researcher evaluating ARTEX, the agent framework that automates intrusions and that VAPOR PANDA has already used. The researcher says the agent escaped its isolated test environment and hit WorkOS unintentionally; the company and an independent forensic firm matched his details against its logs and named him responsible "with high confidence," and he says the data he pulled has been deleted. The agent didn't need anything clever: the app ran on a legacy internal platform, its APIs were exposed to the internet, they accepted a predictable development credential left in production, and there was no rate limiting. It got in on October 25 and nobody noticed until October 30, when an employee spotted unusual usage, by which point it had already downloaded contact details, account histories and summaries of sales calls, support threads and customer Slack conversations. The same week, a Railway employee, on his own time, pointed testing agents at a rival platform, Primcloud, which accused him of trying to break in; he stopped it within 15 minutes, and Railway has banned any testing of outside organizations that isn't done by hand. Lawyers who spoke to Business Insider warn that whoever launches the agent can end up liable. The takeaway: inventory the old internal tools that are visible from the internet, hunt for development keys in production, add rate limits and an alert for abnormal usage, because here it took five days. And if your people test agents, write a closed list of systems they may target: nothing that belongs to someone else. Letting a process run on the UNATTENDED rung starts with knowing where its perimeter ends. Source and Source

OpenAI takes in $11 billion at $1.25 trillion and Anthropic accepts a cheaper IPO: your leverage is in the contract

OpenAI has received $11 billion from a round that values it at $1.25 trillion, according to Bloomberg; the round is still open, could reach about $20 billion and is expected to close in the coming weeks. In September it was seeking $30 billion, and we already covered the cut to $20 billion. Anthropic, meanwhile, is pressing ahead with a U.S. listing before Thanksgiving, on November 26, even if it has to accept a lower valuation, the Financial Times reports; it expects to raise enough that it won't need to sell more equity. Bloomberg puts the valuation prospective buyers are discussing at $1.2 trillion to $1.3 trillion, down from the $1.4 trillion to $1.5 trillion shared with bankers and as much as $2 trillion in early October. There's no confirmed date, price or share count, and Anthropic hasn't commented. It's the next step after the investor doubts we covered on Friday. The takeaway: the two model makers businesses use most are heading into year-end showing revenue to investors and, in Anthropic's case, to public markets every quarter. Our read is that this is your leverage: if you're renewing or expanding before January, ask for a 12-month price lock, usage limits written into the contract and at least 60 days' notice of any change. The Claude Max case shows what happens when those three things aren't in writing. Source and Source

What to watch next

On Monday, November 9, Copilot Chat usage caps reach every organization, and this week will show whether Anthropic kicks off its investor roadshow and sets a price range. In the Claude Max case, watch whether the plaintiff amends his claim and whether Anthropic spells out the weekly cap on its sales page.