noticias ia para empresas agentes de ia ciberseguridad

AI agents now work for the other side too: a single criminal turned them loose on more than 1,200 websites and broke into the sites of more than 100 companies at roughly $50 per break-in, on the same day Microsoft urged companies to treat their own agents like an employee holding the keys. Santander shows the flip side: in its tests, smaller models cut the cost per interaction from about €5 to 3 cents. And money is landing in Spain: Salesforce will invest $1 billion over five years and Linkener has raised €7.5 million to keep watch over companies' energy bills. Picking up from Monday's radar, October 19.

A hacker turns AI agents loose on 1,200 websites and breaks into those of more than 100 companies at $50 apiece: your forgotten website is the way in

Snorre Fagerland, a researcher at cybersecurity firm NorForge, spent two weeks watching a Chinese-speaking criminal hand almost all the work to AI agents. Since September 25, the hacker's agents have tried to break into more than 1,200 e-commerce websites in the United States, Canada, Australia, and Japan, and got into the sites of more than 100 companies, including Deere (the maker of John Deere tractors), SharkNinja and Cummins. The instructions fit in fewer than 100 Chinese characters, each successful intrusion typically took about 23 hours of probing with no human at the controls, and the campaign, which appeared to use a mix of DeepSeek and Claude Code, burned through roughly $5,000 in tokens by Fagerland's estimate: about $50 per site breached. In one case, the agents got a large company's server to run their commands through a flaw in an internet-facing control panel. Deere says it has no indication that sensitive information was accessed, SharkNinja says the hack hit a standalone product support site, and Cummins is looking into a possible attack on a website run by one of its vendors; how much data was taken is unknown. For your company: at $50 a break-in, you don't need to be an interesting target anymore; you just need to have something exposed. Today, list everything you have online (your store, an old campaign site, the test subdomain, the hosting control panel, whatever a vendor runs for you), shut down what you don't use, put two-factor authentication on admin panels and switch on automatic updates in waves, as we suggested for Windows patches on Thursday, October 15. And ask whoever maintains your website, in writing, who updates it and how often. Looking after your websites by hand and from memory means staying MANUAL against an attacker who already runs UNATTENDED. Source

Santander cuts the cost per interaction from €5 to 3 cents with smaller models: the real asset is its test bench

Banco Santander has signed a three-year collaboration framework with Nvidia and laid out what its AI lab has already delivered. The bank is moving from large general models to smaller models specialized in specific banking tasks, and in experimental use cases that approach has cut running costs from about €5 per interaction to about €0.03 "while maintaining good results." In a customer service test, conversation quality improved and costs dropped more than 50%. The key piece is Model Arena, an internal platform that compares open-weight models on real bank tasks to choose the most suitable one for each use case and to test agents before deploying them group-wide. The agreement also spans agents, fraud and anti-money-laundering work, and real-time decisions. Aser Blanco, global head of banking at Nvidia, sums it up: "the cost of each result generated by artificial intelligence is a fundamental metric." These are tests, not the whole bank, and the story comes from the company selling the infrastructure. For your company: you don't need a lab to copy the method. Choose a high-volume repetitive task (order inquiries, classifying emails, extracting data from delivery notes), gather 50 real cases with their correct answers, and run them through the model you use today and a smaller one; track accuracy and cost per resolved case, as we suggested with documents on Friday, October 16. That test suite is what lets you move a process to UNATTENDED with numbers in hand; without it, switching models is a gamble. Source

Salesforce will invest $1 billion in Spain and Linkener raises €7.5 million: the money is going to the people who deploy AI, and to AI that pays for itself

Salesforce will invest $1 billion (about €892 million) in Spain over the next five years. It made the announcement on Monday as it opened its new offices on Madrid's Paseo de la Castellana, its hub for more than 90 countries across Southern Europe, the Middle East, and Africa. The site will house a team of engineers who'll work with Spanish customers to get agents up and running, plus a year-long training program for graduates from across the region. Salesforce has more than 600 employees in Spain and customers such as CaixaBank and Repsol. The same day, Valencia-based Linkener closed a €7.5 million round led by Klima, Alantra's energy transition fund, which counts Enagás among its main investors. Its system, which uses its own remote metering, already digitizes and checks energy bills and issues alerts and savings proposals to more than 2,000 clients, from sole traders to large corporations; the money will go toward proprietary AI models to automate this processing and detect anomalies, and to acquire companies in a highly fragmented sector. Its CEO reckons that keeping consumption under control can save 20% in the services sector. For your company, two takeaways. What's scarce isn't licenses, it's the people who implement them, like the engineers Anthropic wants to train, which we covered on Sunday, October 4: if a partner is selling you AI, ask who will implement it, how many hours they'll put in and what they'll deliver. And the money is flowing to AI tied to a specific bill: before you buy a tool, ask which line of your P&L will show the savings and how they'll be measured. Checking your power bill once a year is MANUAL; having a system check it every month and flag anything odd is one of the first things you can leave UNATTENDED. Source

Microsoft wants you to treat your AI agents like employees with keys: their own account, permissions per task and an expiry date

Microsoft has introduced security features that treat agents as "insiders": people on the inside with access to company systems. Its endpoint monitoring can now spot 60 types of local agents, up from 25, including Claude Code, OpenAI's Codex, and GitHub Copilot; each agent receives its own identity, separate from the person using it, and under the new Dynamic Permissions it starts with no access at all, gets only what its task requires and loses it when the task is done. If the agent drifts, say after reading instructions hidden in a document, the system can keep its read access while blocking writes, require verified human approval or shut it down. As Satya Nadella puts it: "The most trustworthy agentic system will not be the one with the model we trust most. It will be the one designed that enables us to trust the model least." The features are in private preview, with broader availability planned for its Ignite conference on November 17-19, as part of Microsoft 365 E7, which costs $99 per user per month. For your company: the problem we covered on Wednesday, October 14 (that an assistant can read everything the employee using it can read) now has a product answer, but the principle doesn't cost $99. Check which agents are running on your team's computers, give each one its own account or key instead of yours, start with read-only access, and make payments, deletions and emails to customers wait for a human yes. An agent that logs in with your username and your permissions isn't an UNATTENDED process: it's your signature in a machine's hands. Source

What to watch tomorrow

Nvidia's GTC opens in Berlin with Jensen Huang headlining, and it's where the $2 billion European investment that Bloomberg said Nvidia would announce this week could take shape. Also watch whether NorForge publishes its full report on the agent campaign: which flaws were exploited, and whether there are victims beyond the four countries named.