noticias ia para empresas anthropic ciberseguridad

Anthropic has finally shown its third quarter ahead of its IPO: it booked nearly $17 billion in revenue, growth slowed from the spring, it spent more on compute for every dollar it billed, and it disclosed about $600 million in revenue given up to price cuts and customer discounts. Overnight, someone pushed supermarket and butcher shop freezers in seven European countries up to 50°C through their internet-connected controllers. And AI usage is starting to be rationed and audited: ChatGPT and Codex will share a single limit from December, and Larridin raised $30 million to tell companies who is spending what. Catch up on Monday's radar.

Anthropic books $17 billion a quarter and is already giving away $600 million in discounts: get yours in writing

On Monday, Anthropic filed an amended IPO prospectus with the SEC, with numbers through September 30. Third-quarter revenue came in at $16.956 billion, up 1,127% from a year earlier, for $33.242 billion over nine months. Quarter over quarter it grew 47%, after jumping 144% from the first quarter to the second. Compute and infrastructure costs hit $9.863 billion, or 58% of revenue, up from 48% the quarter before, because they include $5 billion of short-term compute that the company says costs "significantly more" than capacity under its multi-year contracts. The result: a $567 million operating loss, after $974 million in operating profit the previous quarter. The $23.291 billion net loss mostly comes from marking its convertible notes to market, not from cash. Two more figures matter to customers: roughly $600 million less revenue in the quarter from price cuts on new model versions and contractual discounts, and $895.4 billion in committed compute spending, half of it due before the end of 2030. More than 11,900 customers now pay it over $100,000 a year. There's still no price range for the shares; that will come in a later amendment, as we noted when the first filing went public. The takeaway: the discounts are real and they're being handed out now, so if you use Claude at volume, ask for yours, and above all lock it in writing with fixed prices and limits for twelve months. A vendor paying more for this quarter's compute, and warning in its own prospectus that it may change usage limits and capabilities, isn't promising you next quarter's discount. An integrated process that depends on it needs that price in the contract, not on a pricing page. Source

Freezers hit 50°C in seven countries: your connected equipment is a way in, too

Early Monday morning, refrigeration systems at supermarkets and shops in Belgium, Luxembourg, Sweden, Ireland, the Netherlands, Estonia and Finland switched their evaporators into defrost mode and drove temperatures as high as 50°C. Around 40 Delhaize stores were hit, along with 15 Carrefour stores in Belgium and 13 SuperValu stores in Ireland, and manufacturer Danfoss says at least 17 businesses in the Netherlands were affected, including butchers and corner shops. A single Delhaize store puts its spoiled stock at €100,000. Belgian installer Sabcobel spent the night disconnecting units from the network and says some need physical modifications, not a software update. The controllers involved are from Danfoss and Copeland and, according to the manufacturers, are mostly discontinued models that no longer get security patches. It's the second wave in three weeks, after October's attack in France hit around 700 victims, and nobody has claimed it. Belgium's cybersecurity agency doesn't know whether AI was used this time, but warns of a huge rise in vulnerabilities being found and exploited with it. And the EU law that will force manufacturers to support products for their expected lifetime doesn't kick in until December 2027. The takeaway: today, list everything you have that's connected and can be controlled remotely, whether that's refrigeration, HVAC, alarms, ovens or compressors. Ask your installer which model it is, whether it still gets updates and how they log in, and insist on a secure connection, never a controller sitting open on the internet. Add an independent temperature alarm that texts your phone. Before you leave any process running unattended, you need to know which machines it runs on and who can touch them. Source · Source

ChatGPT and Codex will share one limit from December: your coders can eat everyone else's allowance

OpenAI posted a help article over the weekend saying that, starting in December, ChatGPT conversations and Codex tasks on the Plus, Pro and Business plans will draw from one shared usage pool, with more demanding models and complex tasks using up more of it. The only exception is Instant chat, the fast mode for simple questions. Today, Codex, ChatGPT Work, ChatGPT for Excel and Workspace Agents already share an allowance for agent tasks, but regular chat had its own. That separation is what goes away. OpenAI hasn't said how big the combined pool will be, only that your remaining balance and reset time will show up in settings. The change comes days after it backed off cutting the $200 Pro plan and doubled limits on other plans, and subscribers on its forum are already saying they won't renew without a usage dashboard broken down by product and model. The takeaway: if the same seats serve people coding with Codex and people asking questions or drafting documents, in December one group can run the other out of capacity by midweek. Use November to measure who uses what, move heavy agent work onto separate accounts, send routine questions to Instant chat, and check the balance in settings during the first week of December. An integrated process can't depend on a quota you can't see. Source

Larridin raises $30 million to audit AI bills: on one team, a single engineer ran up 65% of the spend

Larridin, a San Francisco company founded in 2024 that tracks AI usage, cost and output inside businesses, closed a $30 million Series A led by Pelion Ventures, with Andreessen Horowitz and Gradient Ventures coming back in. That brings its total funding to $47 million, and its customers include SurveyMonkey, Klaviyo and Vertiv. The customer data it reports, which hasn't been independently audited, shows why the money is there: at one customer, 92% of engineers were using AI, but one engineer accounted for about 65% of the team's AI spend in a single week; at another, two tools doing comparable work differed 66-fold in cost per shipped change. Its quarterly report, covering more than 1,000 engineers, adds that AI-assisted code changes rose from 35% in April to 62% in September, while the median wait for a first human review went up 50%. The takeaway: you don't need to buy a platform to get started. Pull each tool's bill by user, divide it by what got delivered, whether that's a quote, a processed invoice or a code change, and check whether the bottleneck has already moved to whoever does the reviewing. Measuring in euros or dollars per deliverable is what separates using AI from knowing whether it pays. Source

What to watch tomorrow

Whether Anthropic files a new amendment with a price range. Whether anyone attributes the attack on the refrigeration systems, and which controller models are confirmed. And whether OpenAI says how big the shared ChatGPT and Codex pool will be.